diff --git a/.gitattributes b/.gitattributes new file mode 100644 index 0000000000000000000000000000000000000000..69a0c28d461e4000c13b9e9c420a0c945cc2568c --- /dev/null +++ b/.gitattributes @@ -0,0 +1,43 @@ +*.7z filter=lfs diff=lfs merge=lfs -text +*.arrow filter=lfs diff=lfs merge=lfs -text +*.bin filter=lfs diff=lfs merge=lfs -text +*.bz2 filter=lfs diff=lfs merge=lfs -text +*.ckpt filter=lfs diff=lfs merge=lfs -text +*.ftz filter=lfs diff=lfs merge=lfs -text +*.gz filter=lfs diff=lfs merge=lfs -text +*.h5 filter=lfs diff=lfs merge=lfs -text +*.joblib filter=lfs diff=lfs merge=lfs -text +*.lfs.* filter=lfs diff=lfs merge=lfs -text +*.mlmodel filter=lfs diff=lfs merge=lfs -text +*.model filter=lfs diff=lfs merge=lfs -text +*.msgpack filter=lfs diff=lfs merge=lfs -text +*.npy filter=lfs diff=lfs merge=lfs -text +*.npz filter=lfs diff=lfs merge=lfs -text +*.onnx filter=lfs diff=lfs merge=lfs -text +*.ot filter=lfs diff=lfs merge=lfs -text +*.parquet filter=lfs diff=lfs merge=lfs -text +*.pb filter=lfs diff=lfs merge=lfs -text +*.pickle filter=lfs diff=lfs merge=lfs -text +*.pkl filter=lfs diff=lfs merge=lfs -text +*.pt filter=lfs diff=lfs merge=lfs -text +*.pth filter=lfs diff=lfs merge=lfs -text +*.rar filter=lfs diff=lfs merge=lfs -text +*.safetensors filter=lfs diff=lfs merge=lfs -text +saved_model/**/* filter=lfs diff=lfs merge=lfs -text +*.tar.* filter=lfs diff=lfs merge=lfs -text +*.tar filter=lfs diff=lfs merge=lfs -text +*.tflite filter=lfs diff=lfs merge=lfs -text +*.tgz filter=lfs diff=lfs merge=lfs -text +*.wasm filter=lfs diff=lfs merge=lfs -text +*.xz filter=lfs diff=lfs merge=lfs -text +*.zip filter=lfs diff=lfs merge=lfs -text +*.zst filter=lfs diff=lfs merge=lfs -text +*tfevents* filter=lfs diff=lfs merge=lfs -text +stage1/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text +tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text +variants/events_brier_seed17/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text +variants/events_ce_seed17/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text +variants/events_paired_seed17/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text +variants/games_api_seed17/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text +variants/games_gold_seed17/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text +variants/local_atomic_seed17/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text diff --git a/GAMES_MODEL_MANIFEST.json b/GAMES_MODEL_MANIFEST.json new file mode 100644 index 0000000000000000000000000000000000000000..43ecc53e51f90015ed32725e9a929b47a967ec8d --- /dev/null +++ b/GAMES_MODEL_MANIFEST.json @@ -0,0 +1,842 @@ +{ + "schema": "nanojev-games-model-release-v1", + "model_repository": "C-Tianyu/NanoJev", + "dataset_repository": "C-Tianyu/NanoJev-Data", + "base_model": "Qwen/Qwen3-0.6B", + "base_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "weight_staging": "hard links to existing frozen weights; configuration and audit files are byte-for-byte copies; no symlinks", + "variants": [ + { + "name": "local_atomic_seed17", + "path": "variants/local_atomic_seed17", + "source_checkpoint": "/home/rwang/nanojev_dev_v4_20260917/runs/local_atomic_seed17", + "weight_sha256": "2b06e4423861f47da7ca53ad9c23a26aebdbfe79de3a238b36c2954ae9c6abf0", + "best_step": 300, + "best_dev_target_ce": 0.48294000757494554, + "training_seconds": 242.69976669549942, + "objective": "gold_distribution", + "loss": "ce", + "data_source_directory": "data/local_maze_v1", + "data_files": [ + { + "source_path": "data/local_maze_v1/train.jsonl", + "sha256": "502f2840f3e15ffac76554f35ef6c2e41c6e25f58f77b5ee3a2f1196cba0c41a", + "bytes": 754633 + }, + { + "source_path": "data/local_maze_v1/dev.jsonl", + "sha256": "bd7f9d4402112770f571210677e329e8673a38caa6bb798b744ee47bc793b9a5", + "bytes": 262359 + }, + { + "source_path": "data/local_maze_v1/calibration.jsonl", + "sha256": "94a233e71eaf3460c939bb0b4b147a0a951931001097070ea7ac8b4b78f39ace", + "bytes": 252234 + }, + { + "source_path": "data/local_maze_v1/test.jsonl", + "sha256": "be7783c395715f9fa2aae7c793c4e5b7c4b1dfe3fcb0880544b577e85b8f3b10", + "bytes": 232377 + }, + { + "source_path": "data/local_maze_v1/ood.jsonl", + "sha256": "e6772998a189880c780b6408c0e348c77324a28204e390767dec54f475d61f4e", + "bytes": 195310 + } + ], + "files": [ + { + "path": "variants/local_atomic_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "2b06e4423861f47da7ca53ad9c23a26aebdbfe79de3a238b36c2954ae9c6abf0" + }, + { + "path": "variants/local_atomic_seed17/config.json", + "bytes": 2498, + "sha256": "3a6b014a77eaa3a79155a9c0c84ce2c005b5a5c66fdf88445067657ac9def176" + }, + { + "path": "variants/local_atomic_seed17/train_log.json", + "bytes": 85584, + "sha256": "5e2a39713c30d714dddd8b7d989aecf1a8105fee24e74b42ef8943a1fd4dff09" + }, + { + "path": "variants/local_atomic_seed17/summary.json", + "bytes": 4341, + "sha256": "3da659f939b0cc8d44614d9258f8660482db548a3f62e6f5c8604f15a9770980" + }, + { + "path": "variants/local_atomic_seed17/target_audit.json", + "bytes": 701867, + "sha256": "ae1cde9eeedf6e84266205fba29b32bcb58489ca7da4f17704d74d2964d6cc79" + }, + { + "path": "variants/local_atomic_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/local_atomic_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/local_atomic_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/local_atomic_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + } + ], + "initial_checkpoint_sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28", + "minimum_inference_patterns": [ + "best.safetensors", + "config.json", + "tokenizer/*", + "backbone_config/*" + ] + }, + { + "name": "games_gold_seed17", + "path": "variants/games_gold_seed17", + "source_checkpoint": "/home/rwang/nanojev_dev_v4_20260917/runs/games_gold_seed17", + "weight_sha256": "5a395c56fad2a9bd4fb664ef78e1e06c19102886d97dd56dfe2676e5896c7e3a", + "best_step": 100, + "best_dev_target_ce": 0.6079203053015539, + "training_seconds": 186.51977808214724, + "objective": "gold_distribution", + "loss": "ce", + "data_source_directory": "data/scaled_games_v4b/policy", + "data_files": [ + { + "source_path": "data/scaled_games_v4b/policy/train.jsonl", + "sha256": "af6065315d383bce2f137b3095cc9912aafe1edf0a3ee7c119c9b7af4a0ac399", + "bytes": 1776280 + }, + { + "source_path": "data/scaled_games_v4b/policy/dev.jsonl", + "sha256": "5e34f86e60dffce37d9454c4ffb978c987ba8467de35c5f79dab99bdb24f3d6b", + "bytes": 601279 + }, + { + "source_path": "data/scaled_games_v4b/policy/calibration.jsonl", + "sha256": "c2b279ecd835068c20bc328c34bca970feee1e8afab67b1f7dea0c201f0d824f", + "bytes": 584303 + }, + { + "source_path": "data/scaled_games_v4b/policy/test.jsonl", + "sha256": "54e573515e5d1e2c55308763a857d9bb46f2c02b2cbd23e50a4544587a69501c", + "bytes": 569530 + }, + { + "source_path": "data/scaled_games_v4b/policy/ood.jsonl", + "sha256": "ac4ea92590a95f40cd2134ad64f450c369601745b36d91ebf5692d9f39ea3135", + "bytes": 375280 + } + ], + "files": [ + { + "path": "variants/games_gold_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "5a395c56fad2a9bd4fb664ef78e1e06c19102886d97dd56dfe2676e5896c7e3a" + }, + { + "path": "variants/games_gold_seed17/config.json", + "bytes": 2562, + "sha256": "965c5905315315ff591d6c2eb581146b749104e65afdd463e3a3118f70baaaff" + }, + { + "path": "variants/games_gold_seed17/train_log.json", + "bytes": 42887, + "sha256": "a1c2de04bf59f1da31ab4255be69a5b5960fde183b8a459d2cd8b1f7c618ae28" + }, + { + "path": "variants/games_gold_seed17/summary.json", + "bytes": 4530, + "sha256": "f7de836505d40a4f779258155455fc1d3d8194e468cae269bcb942c13441ff5c" + }, + { + "path": "variants/games_gold_seed17/target_audit.json", + "bytes": 1831844, + "sha256": "62e50d044875e9914b7edce5d1321c873eb923a31fefaf0a32d8a13e0e755b1b" + }, + { + "path": "variants/games_gold_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/games_gold_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/games_gold_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/games_gold_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + } + ], + "initial_checkpoint_sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28", + "minimum_inference_patterns": [ + "best.safetensors", + "config.json", + "tokenizer/*", + "backbone_config/*" + ] + }, + { + "name": "games_api_seed17", + "path": "variants/games_api_seed17", + "source_checkpoint": "/home/rwang/nanojev_dev_v4_20260917/runs/games_api_seed17", + "weight_sha256": "45829fc896391847220b67dcf6fc0702427d2f184be384649bcf9343c1552477", + "best_step": 150, + "best_dev_target_ce": 0.7130707709730839, + "training_seconds": 182.26331992074847, + "objective": "teacher", + "loss": "ce", + "data_source_directory": "data/scaled_games_labeled_v4", + "data_files": [ + { + "source_path": "data/scaled_games_labeled_v4/train.jsonl", + "sha256": "e21d05f1c34253687f4e7b78974503e7c05a7e145e4cbfc306246938d9f6317c", + "bytes": 1542530 + }, + { + "source_path": "data/scaled_games_labeled_v4/dev.jsonl", + "sha256": "fd7780fd6bf94263f5653cea8b3575925447981602dba61d0c7ff96b8433ea88", + "bytes": 524030 + }, + { + "source_path": "data/scaled_games_labeled_v4/calibration.jsonl", + "sha256": "9632b050ee1069c4393e275fc878c2681f9c36a501c4c6584df19b3fb88d76fd", + "bytes": 507336 + }, + { + "source_path": "data/scaled_games_labeled_v4/test.jsonl", + "sha256": "82b62b39f56bcd703e6e7c911aafb20203ff7b48ebf4f4f4c3444694e4823fa0", + "bytes": 494657 + }, + { + "source_path": "data/scaled_games_labeled_v4/ood.jsonl", + "sha256": "568c11c100126b3848fb29c9cd3d4b8dc22be42fe2df43a3bb3cf8e7b0128ae1", + "bytes": 348817 + } + ], + "files": [ + { + "path": "variants/games_api_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "45829fc896391847220b67dcf6fc0702427d2f184be384649bcf9343c1552477" + }, + { + "path": "variants/games_api_seed17/config.json", + "bytes": 2693, + "sha256": "509c449f3088e167f95ea284370763489ca182e787d7765efbcffc377b767931" + }, + { + "path": "variants/games_api_seed17/train_log.json", + "bytes": 42439, + "sha256": "aca1693ee0a2c1c1c574416329e28a089adaa93014a071b09ac4350b62e12050" + }, + { + "path": "variants/games_api_seed17/summary.json", + "bytes": 3996, + "sha256": "304fde47ce4cef20bd06ed088c55f4a8bc2c8da3f68e3f5a9520a5bdbf14dbdb" + }, + { + "path": "variants/games_api_seed17/target_audit.json", + "bytes": 2012764, + "sha256": "28a1666127b407bbfdaf8bcf99ca386c66a814a37ef78221bda12cdce9f8ab38" + }, + { + "path": "variants/games_api_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/games_api_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/games_api_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/games_api_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + } + ], + "initial_checkpoint_sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28", + "minimum_inference_patterns": [ + "best.safetensors", + "config.json", + "tokenizer/*", + "backbone_config/*" + ] + }, + { + "name": "events_ce_seed17", + "path": "variants/events_ce_seed17", + "source_checkpoint": "/home/rwang/nanojev_dev_v4_20260917/runs/events_ce_seed17", + "weight_sha256": "582c01139e29807b5f2690335271cf5647583cbfc3dd963d5967792bb3963992", + "best_step": 100, + "best_dev_target_ce": 0.47277030621569244, + "training_seconds": 96.74807532411069, + "objective": "observed_outcome", + "loss": "ce", + "data_source_directory": "data/scaled_games_v4/events", + "data_files": [ + { + "source_path": "data/scaled_games_v4/events/train.jsonl", + "sha256": "4cebf698e4b71dc37d4d0e985b6577dd565ac370ac82b6f47f375aa6af69f8c9", + "bytes": 4061794 + }, + { + "source_path": "data/scaled_games_v4/events/dev.jsonl", + "sha256": "0c352ee1064a5fb5d2aacad59cbea57afb897886d3d09e5ca108c866a5f893db", + "bytes": 1403407 + }, + { + "source_path": "data/scaled_games_v4/events/calibration.jsonl", + "sha256": "4a8a8a2b81c3c46df6bb4c8690e97374dd0165cbef40566bb7f87c1f425d5562", + "bytes": 1345231 + }, + { + "source_path": "data/scaled_games_v4/events/test.jsonl", + "sha256": "884bec71891668199925809d8708347f56f7bf06e8363e4a436347967b1423e3", + "bytes": 1299634 + }, + { + "source_path": "data/scaled_games_v4/events/ood.jsonl", + "sha256": "3c00d845524634ca3b34559761cdaf0c52b7b7633d5517d1e7ee7f31697cea48", + "bytes": 1154862 + } + ], + "files": [ + { + "path": "variants/events_ce_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "582c01139e29807b5f2690335271cf5647583cbfc3dd963d5967792bb3963992" + }, + { + "path": "variants/events_ce_seed17/config.json", + "bytes": 2555, + "sha256": "52f0309ceeefe51d8cadda95b6f0985978b2d523d758194f682932892d82a566" + }, + { + "path": "variants/events_ce_seed17/train_log.json", + "bytes": 28195, + "sha256": "c5e20da3ea98546ac71a9d2a264b53d52bcfb3387a7cd5a57d65c7cc03c9623e" + }, + { + "path": "variants/events_ce_seed17/summary.json", + "bytes": 3840, + "sha256": "b5f1a6592da085ea88ca7676fbaf11b39874b097cdc70a3424c992010b27c3a5" + }, + { + "path": "variants/events_ce_seed17/target_audit.json", + "bytes": 1334523, + "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181" + }, + { + "path": "variants/events_ce_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/events_ce_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/events_ce_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/events_ce_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + } + ], + "initial_checkpoint_sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28", + "minimum_inference_patterns": [ + "best.safetensors", + "config.json", + "tokenizer/*", + "backbone_config/*" + ] + }, + { + "name": "events_brier_seed17", + "path": "variants/events_brier_seed17", + "source_checkpoint": "/home/rwang/nanojev_dev_v4_20260917/runs/events_brier_seed17", + "weight_sha256": "0a8d8b14c97743eed4310d461a7ae5cff0400cb87eaa6dc1f8c5dc3c59988651", + "best_step": 100, + "best_dev_target_ce": 0.4421807345122086, + "training_seconds": 94.21064267028123, + "objective": "observed_outcome", + "loss": "brier", + "data_source_directory": "data/scaled_games_v4/events", + "data_files": [ + { + "source_path": "data/scaled_games_v4/events/train.jsonl", + "sha256": "4cebf698e4b71dc37d4d0e985b6577dd565ac370ac82b6f47f375aa6af69f8c9", + "bytes": 4061794 + }, + { + "source_path": "data/scaled_games_v4/events/dev.jsonl", + "sha256": "0c352ee1064a5fb5d2aacad59cbea57afb897886d3d09e5ca108c866a5f893db", + "bytes": 1403407 + }, + { + "source_path": "data/scaled_games_v4/events/calibration.jsonl", + "sha256": "4a8a8a2b81c3c46df6bb4c8690e97374dd0165cbef40566bb7f87c1f425d5562", + "bytes": 1345231 + }, + { + "source_path": "data/scaled_games_v4/events/test.jsonl", + "sha256": "884bec71891668199925809d8708347f56f7bf06e8363e4a436347967b1423e3", + "bytes": 1299634 + }, + { + "source_path": "data/scaled_games_v4/events/ood.jsonl", + "sha256": "3c00d845524634ca3b34559761cdaf0c52b7b7633d5517d1e7ee7f31697cea48", + "bytes": 1154862 + } + ], + "files": [ + { + "path": "variants/events_brier_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "0a8d8b14c97743eed4310d461a7ae5cff0400cb87eaa6dc1f8c5dc3c59988651" + }, + { + "path": "variants/events_brier_seed17/config.json", + "bytes": 2561, + "sha256": "bca3a9b31a8f528bc144433f1b2f4f9f27f78f50ceb5957c99949e246c850ac2" + }, + { + "path": "variants/events_brier_seed17/train_log.json", + "bytes": 28219, + "sha256": "e09fdb39c212589b56af8f5cbae6f51fa1d1f91de73a63636e8661a02f2312f9" + }, + { + "path": "variants/events_brier_seed17/summary.json", + "bytes": 3829, + "sha256": "daaf0e3739fb244368662944e26b939278a31beac7c1c538ce3e72aa70536def" + }, + { + "path": "variants/events_brier_seed17/target_audit.json", + "bytes": 1334523, + "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181" + }, + { + "path": "variants/events_brier_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/events_brier_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/events_brier_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/events_brier_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + } + ], + "initial_checkpoint_sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28", + "minimum_inference_patterns": [ + "best.safetensors", + "config.json", + "tokenizer/*", + "backbone_config/*" + ] + }, + { + "name": "events_paired_seed17", + "path": "variants/events_paired_seed17", + "source_checkpoint": "/home/rwang/nanojev_dev_v4_20260917/runs/events_paired_seed17", + "weight_sha256": "bbefcf829b9e36253ef46ab076f7d8fca00ffb33762e8358dbe780b3605bb777", + "best_step": 100, + "best_dev_target_ce": 0.45392147839038, + "training_seconds": 100.81704722624272, + "objective": "observed_outcome", + "loss": "paired_brier_pg", + "data_source_directory": "data/scaled_games_v4/events", + "data_files": [ + { + "source_path": "data/scaled_games_v4/events/train.jsonl", + "sha256": "4cebf698e4b71dc37d4d0e985b6577dd565ac370ac82b6f47f375aa6af69f8c9", + "bytes": 4061794 + }, + { + "source_path": "data/scaled_games_v4/events/dev.jsonl", + "sha256": "0c352ee1064a5fb5d2aacad59cbea57afb897886d3d09e5ca108c866a5f893db", + "bytes": 1403407 + }, + { + "source_path": "data/scaled_games_v4/events/calibration.jsonl", + "sha256": "4a8a8a2b81c3c46df6bb4c8690e97374dd0165cbef40566bb7f87c1f425d5562", + "bytes": 1345231 + }, + { + "source_path": "data/scaled_games_v4/events/test.jsonl", + "sha256": "884bec71891668199925809d8708347f56f7bf06e8363e4a436347967b1423e3", + "bytes": 1299634 + }, + { + "source_path": "data/scaled_games_v4/events/ood.jsonl", + "sha256": "3c00d845524634ca3b34559761cdaf0c52b7b7633d5517d1e7ee7f31697cea48", + "bytes": 1154862 + } + ], + "files": [ + { + "path": "variants/events_paired_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "bbefcf829b9e36253ef46ab076f7d8fca00ffb33762e8358dbe780b3605bb777" + }, + { + "path": "variants/events_paired_seed17/config.json", + "bytes": 2671, + "sha256": "9bdffa977eb28e4692ba19f53b05a7ad9afd9f37665848f9db24de4e829f1acb" + }, + { + "path": "variants/events_paired_seed17/train_log.json", + "bytes": 28356, + "sha256": "33f337e289dab3afce8199fc7b72a6c766634986f49276ba8581ea2210b673cf" + }, + { + "path": "variants/events_paired_seed17/summary.json", + "bytes": 3817, + "sha256": "82de7efe05e55bedc75e5b50b5b9ca867aa1f100cb4c4430b9b67524c6e82de1" + }, + { + "path": "variants/events_paired_seed17/target_audit.json", + "bytes": 1334523, + "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181" + }, + { + "path": "variants/events_paired_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/events_paired_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/events_paired_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/events_paired_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + } + ], + "initial_checkpoint_sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28", + "minimum_inference_patterns": [ + "best.safetensors", + "config.json", + "tokenizer/*", + "backbone_config/*" + ] + } + ], + "files": [ + { + "path": "variants/local_atomic_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "2b06e4423861f47da7ca53ad9c23a26aebdbfe79de3a238b36c2954ae9c6abf0" + }, + { + "path": "variants/local_atomic_seed17/config.json", + "bytes": 2498, + "sha256": "3a6b014a77eaa3a79155a9c0c84ce2c005b5a5c66fdf88445067657ac9def176" + }, + { + "path": "variants/local_atomic_seed17/train_log.json", + "bytes": 85584, + "sha256": "5e2a39713c30d714dddd8b7d989aecf1a8105fee24e74b42ef8943a1fd4dff09" + }, + { + "path": "variants/local_atomic_seed17/summary.json", + "bytes": 4341, + "sha256": "3da659f939b0cc8d44614d9258f8660482db548a3f62e6f5c8604f15a9770980" + }, + { + "path": "variants/local_atomic_seed17/target_audit.json", + "bytes": 701867, + "sha256": "ae1cde9eeedf6e84266205fba29b32bcb58489ca7da4f17704d74d2964d6cc79" + }, + { + "path": "variants/local_atomic_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/local_atomic_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/local_atomic_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/local_atomic_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + }, + { + "path": "variants/games_gold_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "5a395c56fad2a9bd4fb664ef78e1e06c19102886d97dd56dfe2676e5896c7e3a" + }, + { + "path": "variants/games_gold_seed17/config.json", + "bytes": 2562, + "sha256": "965c5905315315ff591d6c2eb581146b749104e65afdd463e3a3118f70baaaff" + }, + { + "path": "variants/games_gold_seed17/train_log.json", + "bytes": 42887, + "sha256": "a1c2de04bf59f1da31ab4255be69a5b5960fde183b8a459d2cd8b1f7c618ae28" + }, + { + "path": "variants/games_gold_seed17/summary.json", + "bytes": 4530, + "sha256": "f7de836505d40a4f779258155455fc1d3d8194e468cae269bcb942c13441ff5c" + }, + { + "path": "variants/games_gold_seed17/target_audit.json", + "bytes": 1831844, + "sha256": "62e50d044875e9914b7edce5d1321c873eb923a31fefaf0a32d8a13e0e755b1b" + }, + { + "path": "variants/games_gold_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/games_gold_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/games_gold_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/games_gold_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + }, + { + "path": "variants/games_api_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "45829fc896391847220b67dcf6fc0702427d2f184be384649bcf9343c1552477" + }, + { + "path": "variants/games_api_seed17/config.json", + "bytes": 2693, + "sha256": "509c449f3088e167f95ea284370763489ca182e787d7765efbcffc377b767931" + }, + { + "path": "variants/games_api_seed17/train_log.json", + "bytes": 42439, + "sha256": "aca1693ee0a2c1c1c574416329e28a089adaa93014a071b09ac4350b62e12050" + }, + { + "path": "variants/games_api_seed17/summary.json", + "bytes": 3996, + "sha256": "304fde47ce4cef20bd06ed088c55f4a8bc2c8da3f68e3f5a9520a5bdbf14dbdb" + }, + { + "path": "variants/games_api_seed17/target_audit.json", + "bytes": 2012764, + "sha256": "28a1666127b407bbfdaf8bcf99ca386c66a814a37ef78221bda12cdce9f8ab38" + }, + { + "path": "variants/games_api_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/games_api_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/games_api_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/games_api_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + }, + { + "path": "variants/events_ce_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "582c01139e29807b5f2690335271cf5647583cbfc3dd963d5967792bb3963992" + }, + { + "path": "variants/events_ce_seed17/config.json", + "bytes": 2555, + "sha256": "52f0309ceeefe51d8cadda95b6f0985978b2d523d758194f682932892d82a566" + }, + { + "path": "variants/events_ce_seed17/train_log.json", + "bytes": 28195, + "sha256": "c5e20da3ea98546ac71a9d2a264b53d52bcfb3387a7cd5a57d65c7cc03c9623e" + }, + { + "path": "variants/events_ce_seed17/summary.json", + "bytes": 3840, + "sha256": "b5f1a6592da085ea88ca7676fbaf11b39874b097cdc70a3424c992010b27c3a5" + }, + { + "path": "variants/events_ce_seed17/target_audit.json", + "bytes": 1334523, + "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181" + }, + { + "path": "variants/events_ce_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/events_ce_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/events_ce_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/events_ce_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + }, + { + "path": "variants/events_brier_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "0a8d8b14c97743eed4310d461a7ae5cff0400cb87eaa6dc1f8c5dc3c59988651" + }, + { + "path": "variants/events_brier_seed17/config.json", + "bytes": 2561, + "sha256": "bca3a9b31a8f528bc144433f1b2f4f9f27f78f50ceb5957c99949e246c850ac2" + }, + { + "path": "variants/events_brier_seed17/train_log.json", + "bytes": 28219, + "sha256": "e09fdb39c212589b56af8f5cbae6f51fa1d1f91de73a63636e8661a02f2312f9" + }, + { + "path": "variants/events_brier_seed17/summary.json", + "bytes": 3829, + "sha256": "daaf0e3739fb244368662944e26b939278a31beac7c1c538ce3e72aa70536def" + }, + { + "path": "variants/events_brier_seed17/target_audit.json", + "bytes": 1334523, + "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181" + }, + { + "path": "variants/events_brier_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/events_brier_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/events_brier_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/events_brier_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + }, + { + "path": "variants/events_paired_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "bbefcf829b9e36253ef46ab076f7d8fca00ffb33762e8358dbe780b3605bb777" + }, + { + "path": "variants/events_paired_seed17/config.json", + "bytes": 2671, + "sha256": "9bdffa977eb28e4692ba19f53b05a7ad9afd9f37665848f9db24de4e829f1acb" + }, + { + "path": "variants/events_paired_seed17/train_log.json", + "bytes": 28356, + "sha256": "33f337e289dab3afce8199fc7b72a6c766634986f49276ba8581ea2210b673cf" + }, + { + "path": "variants/events_paired_seed17/summary.json", + "bytes": 3817, + "sha256": "82de7efe05e55bedc75e5b50b5b9ca867aa1f100cb4c4430b9b67524c6e82de1" + }, + { + "path": "variants/events_paired_seed17/target_audit.json", + "bytes": 1334523, + "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181" + }, + { + "path": "variants/events_paired_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/events_paired_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/events_paired_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/events_paired_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + } + ], + "total_variant_files": 54, + "total_variant_bytes": 14387654859 +} diff --git a/MODEL_MANIFEST.json b/MODEL_MANIFEST.json new file mode 100644 index 0000000000000000000000000000000000000000..f1c3d02d5671fd40eaba275df2c60ee7ecab35ac --- /dev/null +++ b/MODEL_MANIFEST.json @@ -0,0 +1,383 @@ +{ + "schema_version": "nanojev-model-release-v1", + "model_id": "C-Tianyu/NanoJev", + "base_model": "Qwen/Qwen3-0.6B", + "base_revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "dataset": "C-Tianyu/NanoJev-Data", + "files": [ + { + "path": "README.md", + "bytes": 8030, + "sha256": "6f76023d0ad20ba48f501506a48d2b5df2136e22d8e73c1cc981e16295499983" + }, + { + "path": "backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + }, + { + "path": "best.safetensors", + "bytes": 2385039280, + "sha256": "fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28" + }, + { + "path": "config.json", + "bytes": 2332, + "sha256": "8139aff38992e200c92de038169b1924412dc7a73c72fc90d86d3db4734fc6c1" + }, + { + "path": "stage1/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + }, + { + "path": "stage1/best.safetensors", + "bytes": 2385039280, + "sha256": "231b5178098477d9f82ae2cf38786d9ab8e0a80a019f009e55b772a66882cec4" + }, + { + "path": "stage1/config.json", + "bytes": 2229, + "sha256": "eb010e7a9a4dc3cc46a9fc63c3c516a495667c81697c4b80d54059e91fc69cca" + }, + { + "path": "stage1/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "stage1/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "stage1/tokenizer/tokenizer_config.json", + "bytes": 693, + "sha256": "c4f4c62b741ab2940841480f9942f378d7d3878d47c2babb862af7381c59b952" + }, + { + "path": "tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/local_atomic_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "2b06e4423861f47da7ca53ad9c23a26aebdbfe79de3a238b36c2954ae9c6abf0" + }, + { + "path": "variants/local_atomic_seed17/config.json", + "bytes": 2498, + "sha256": "3a6b014a77eaa3a79155a9c0c84ce2c005b5a5c66fdf88445067657ac9def176" + }, + { + "path": "variants/local_atomic_seed17/train_log.json", + "bytes": 85584, + "sha256": "5e2a39713c30d714dddd8b7d989aecf1a8105fee24e74b42ef8943a1fd4dff09" + }, + { + "path": "variants/local_atomic_seed17/summary.json", + "bytes": 4341, + "sha256": "3da659f939b0cc8d44614d9258f8660482db548a3f62e6f5c8604f15a9770980" + }, + { + "path": "variants/local_atomic_seed17/target_audit.json", + "bytes": 701867, + "sha256": "ae1cde9eeedf6e84266205fba29b32bcb58489ca7da4f17704d74d2964d6cc79" + }, + { + "path": "variants/local_atomic_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/local_atomic_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/local_atomic_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/local_atomic_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + }, + { + "path": "variants/games_gold_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "5a395c56fad2a9bd4fb664ef78e1e06c19102886d97dd56dfe2676e5896c7e3a" + }, + { + "path": "variants/games_gold_seed17/config.json", + "bytes": 2562, + "sha256": "965c5905315315ff591d6c2eb581146b749104e65afdd463e3a3118f70baaaff" + }, + { + "path": "variants/games_gold_seed17/train_log.json", + "bytes": 42887, + "sha256": "a1c2de04bf59f1da31ab4255be69a5b5960fde183b8a459d2cd8b1f7c618ae28" + }, + { + "path": "variants/games_gold_seed17/summary.json", + "bytes": 4530, + "sha256": "f7de836505d40a4f779258155455fc1d3d8194e468cae269bcb942c13441ff5c" + }, + { + "path": "variants/games_gold_seed17/target_audit.json", + "bytes": 1831844, + "sha256": "62e50d044875e9914b7edce5d1321c873eb923a31fefaf0a32d8a13e0e755b1b" + }, + { + "path": "variants/games_gold_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/games_gold_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/games_gold_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/games_gold_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + }, + { + "path": "variants/games_api_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "45829fc896391847220b67dcf6fc0702427d2f184be384649bcf9343c1552477" + }, + { + "path": "variants/games_api_seed17/config.json", + "bytes": 2693, + "sha256": "509c449f3088e167f95ea284370763489ca182e787d7765efbcffc377b767931" + }, + { + "path": "variants/games_api_seed17/train_log.json", + "bytes": 42439, + "sha256": "aca1693ee0a2c1c1c574416329e28a089adaa93014a071b09ac4350b62e12050" + }, + { + "path": "variants/games_api_seed17/summary.json", + "bytes": 3996, + "sha256": "304fde47ce4cef20bd06ed088c55f4a8bc2c8da3f68e3f5a9520a5bdbf14dbdb" + }, + { + "path": "variants/games_api_seed17/target_audit.json", + "bytes": 2012764, + "sha256": "28a1666127b407bbfdaf8bcf99ca386c66a814a37ef78221bda12cdce9f8ab38" + }, + { + "path": "variants/games_api_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/games_api_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/games_api_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/games_api_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + }, + { + "path": "variants/events_ce_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "582c01139e29807b5f2690335271cf5647583cbfc3dd963d5967792bb3963992" + }, + { + "path": "variants/events_ce_seed17/config.json", + "bytes": 2555, + "sha256": "52f0309ceeefe51d8cadda95b6f0985978b2d523d758194f682932892d82a566" + }, + { + "path": "variants/events_ce_seed17/train_log.json", + "bytes": 28195, + "sha256": "c5e20da3ea98546ac71a9d2a264b53d52bcfb3387a7cd5a57d65c7cc03c9623e" + }, + { + "path": "variants/events_ce_seed17/summary.json", + "bytes": 3840, + "sha256": "b5f1a6592da085ea88ca7676fbaf11b39874b097cdc70a3424c992010b27c3a5" + }, + { + "path": "variants/events_ce_seed17/target_audit.json", + "bytes": 1334523, + "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181" + }, + { + "path": "variants/events_ce_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/events_ce_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/events_ce_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/events_ce_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + }, + { + "path": "variants/events_brier_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "0a8d8b14c97743eed4310d461a7ae5cff0400cb87eaa6dc1f8c5dc3c59988651" + }, + { + "path": "variants/events_brier_seed17/config.json", + "bytes": 2561, + "sha256": "bca3a9b31a8f528bc144433f1b2f4f9f27f78f50ceb5957c99949e246c850ac2" + }, + { + "path": "variants/events_brier_seed17/train_log.json", + "bytes": 28219, + "sha256": "e09fdb39c212589b56af8f5cbae6f51fa1d1f91de73a63636e8661a02f2312f9" + }, + { + "path": "variants/events_brier_seed17/summary.json", + "bytes": 3829, + "sha256": "daaf0e3739fb244368662944e26b939278a31beac7c1c538ce3e72aa70536def" + }, + { + "path": "variants/events_brier_seed17/target_audit.json", + "bytes": 1334523, + "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181" + }, + { + "path": "variants/events_brier_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/events_brier_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/events_brier_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/events_brier_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + }, + { + "path": "variants/events_paired_seed17/best.safetensors", + "bytes": 2385039280, + "sha256": "bbefcf829b9e36253ef46ab076f7d8fca00ffb33762e8358dbe780b3605bb777" + }, + { + "path": "variants/events_paired_seed17/config.json", + "bytes": 2671, + "sha256": "9bdffa977eb28e4692ba19f53b05a7ad9afd9f37665848f9db24de4e829f1acb" + }, + { + "path": "variants/events_paired_seed17/train_log.json", + "bytes": 28356, + "sha256": "33f337e289dab3afce8199fc7b72a6c766634986f49276ba8581ea2210b673cf" + }, + { + "path": "variants/events_paired_seed17/summary.json", + "bytes": 3817, + "sha256": "82de7efe05e55bedc75e5b50b5b9ca867aa1f100cb4c4430b9b67524c6e82de1" + }, + { + "path": "variants/events_paired_seed17/target_audit.json", + "bytes": 1334523, + "sha256": "7a9b2635cf87607e9f2a47fa1c2e037f1e5b13c18bcbbc5898cff09796222181" + }, + { + "path": "variants/events_paired_seed17/tokenizer/chat_template.jinja", + "bytes": 4168, + "sha256": "a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8" + }, + { + "path": "variants/events_paired_seed17/tokenizer/tokenizer.json", + "bytes": 11422650, + "sha256": "be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506" + }, + { + "path": "variants/events_paired_seed17/tokenizer/tokenizer_config.json", + "bytes": 692, + "sha256": "1cc816812993bff176eb4f7495433b736f06fba9b6e7b05cac7b4a1780650c95" + }, + { + "path": "variants/events_paired_seed17/backbone_config/config.json", + "bytes": 1417, + "sha256": "30c011854471509747858ac07ffb8a4dab2bc6c04035b184729b69500c557c54" + }, + { + "path": "GAMES_MODEL_MANIFEST.json", + "bytes": 32465, + "sha256": "ed8180264c7052c7806c11b6f2d6606b673aa41eb1d636597393ec89843b94f9" + } + ], + "stage1_input_sha256": "7294765b80e751fc5aee7aba906b28a8ea80d6f147253d3f6c3e4f491de0b2d7", + "stage2_input_sha256": "3a638666206e8ffacadd8beeaf0246687e546716deb8b06adf5e23a228890cdc", + "game_variants": [ + { + "name": "local_atomic_seed17", + "path": "variants/local_atomic_seed17", + "weight_sha256": "2b06e4423861f47da7ca53ad9c23a26aebdbfe79de3a238b36c2954ae9c6abf0" + }, + { + "name": "games_gold_seed17", + "path": "variants/games_gold_seed17", + "weight_sha256": "5a395c56fad2a9bd4fb664ef78e1e06c19102886d97dd56dfe2676e5896c7e3a" + }, + { + "name": "games_api_seed17", + "path": "variants/games_api_seed17", + "weight_sha256": "45829fc896391847220b67dcf6fc0702427d2f184be384649bcf9343c1552477" + }, + { + "name": "events_ce_seed17", + "path": "variants/events_ce_seed17", + "weight_sha256": "582c01139e29807b5f2690335271cf5647583cbfc3dd963d5967792bb3963992" + }, + { + "name": "events_brier_seed17", + "path": "variants/events_brier_seed17", + "weight_sha256": "0a8d8b14c97743eed4310d461a7ae5cff0400cb87eaa6dc1f8c5dc3c59988651" + }, + { + "name": "events_paired_seed17", + "path": "variants/events_paired_seed17", + "weight_sha256": "bbefcf829b9e36253ef46ab076f7d8fca00ffb33762e8358dbe780b3605bb777" + } + ] +} diff --git a/README.md b/README.md new file mode 100644 index 0000000000000000000000000000000000000000..132b770d8e382ef1bd20e6aba1d37e574048d07d --- /dev/null +++ b/README.md @@ -0,0 +1,224 @@ +--- +language: +- en +- zh +base_model: Qwen/Qwen3-0.6B +library_name: pytorch +tags: +- nanojev +- decision-model +- probability-distributions +- parallel-inference +- navigation +--- + +# NanoJev — A nano replica of [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) + +**A 0.6B parallel decision model that returns complete probability distributions over dynamic candidates.** + +Give NanoJev multiple states and questions in one request. It evaluates the candidate paths in one batched backbone forward pass and returns structured decisions with **zero output-token decoding**. + +[Source code](https://github.com/TianyuCodings/NanoJev) · [Dataset](https://huggingface.co/datasets/C-Tianyu/NanoJev-Data) · [Pipeline commands](https://github.com/TianyuCodings/NanoJev/blob/main/research/pipeline_runbook.md) + +## Features + +- **Qwen3-0.6B backbone** with structured decision heads. +- **Multiple states and questions** evaluated in one batched forward. +- **Choice:** 2–255 dynamic candidates with a probability for each candidate. +- **Boolean:** the probability that a proposition is true. +- **Score:** 2–10 ordered levels, their complete distribution, and an expected score. +- **Full distributions** for greedy selection or probability sampling. +- **Persistent serving:** load a checkpoint once and reuse it across requests. + +Measured in the running service: **6 states · 18 questions · 44 candidate paths · 1 backbone forward**. + +## Maze, Snake, and calibrated-decision checkpoints + +The game showcase uses two specialized checkpoints. Every variant below includes complete weights, configuration, tokenizer, and backbone configuration, plus its training summary. + +| Variant directory | Use | +|---|---| +| `variants/local_atomic_seed17` | Local safety judgments used in the 50×50 maze showcase | +| `variants/games_gold_seed17` | Dynamic action choices used in the 12×12 Snake showcase | +| `variants/games_api_seed17` | Full-map game-question comparison | +| `variants/events_ce_seed17` | Observed-event cross-entropy control | +| `variants/events_brier_seed17` | Observed-event Brier control | +| `variants/events_paired_seed17` | RLCD-inspired paired proper-reward experiment | + +[Watch the game demos](https://github.com/TianyuCodings/NanoJev#recorded-showcase-runs) · [Game data](https://huggingface.co/datasets/C-Tianyu/NanoJev-Data/tree/main/games_v4) · [Variant manifest](GAMES_MODEL_MANIFEST.json) + +The recorded 50×50 maze run reaches its goal in 244 attempts. The selected 12×12 Snake run collects 27 food and survives 256 steps with greedy control. These are model decisions composed with the shared planning code described in the repository. + +Download the exact model for either showcase: + +```python +from pathlib import Path +from huggingface_hub import snapshot_download + +variant = "local_atomic_seed17" # Use "games_gold_seed17" for Snake. +snapshot = snapshot_download( + repo_id="C-Tianyu/NanoJev", + allow_patterns=[f"variants/{variant}/*"], +) +checkpoint_dir = Path(snapshot) / "variants" / variant +print(checkpoint_dir) +``` + +Pass the printed directory to the matching evaluator's `--checkpoint` argument, or the persistent service's `--checkpoint-dir`. Full commands are in the [game release guide](https://github.com/TianyuCodings/NanoJev/blob/main/docs/GAME_RELEASE.md). + +## Earlier navigation performance + +**Controller: T=1 probability sampling.** The complete benchmark contains 20 test maps and 20 OOD maps. + +| System | 4×4 test | 6×6 OOD | +|---|---:|---:| +| **trained NanoJev** | **19/20 — 95%** | **18/20 — 90%** | +| Jev | 20/20 — 100% | 19/20 — 95% | +| Original Qwen3-0.6B | 7/20 — 35% | 3/20 — 15% | + +Original Qwen is pretrained without task-specific fine-tuning. Its action distribution is obtained from the native language-model head, conditioned on the offered A–D answer tokens. + +## Download the base release checkpoint + +The base release checkpoint for the earlier navigation benchmark is stored at the repository root: + +```text +best.safetensors +config.json +tokenizer/ +backbone_config/ +``` + +Use `snapshot_download` to retrieve only the final checkpoint files: + +```python +from huggingface_hub import snapshot_download + +checkpoint_dir = snapshot_download( + repo_id="C-Tianyu/NanoJev", + allow_patterns=[ + "best.safetensors", + "config.json", + "tokenizer/*", + "backbone_config/*", + ], +) +print(checkpoint_dir) +``` + +Use the printed snapshot directory directly as `--checkpoint-dir`. + +## Run inference + +Clone the implementation and install its recorded Python dependencies in a compatible CUDA environment: + +```bash +git clone https://github.com/TianyuCodings/NanoJev.git +cd NanoJev +python -m pip install -r requirements-toy.txt +``` + +Save this request as `request.json`: + +```json +{ + "states": [{ + "id": "living_room", + "state": "The room is 29 degrees, the target is 24 degrees, and someone is home.", + "questions": { + "action": { + "type": "choice", + "instructions": "Choose the action that most directly lowers the room temperature.", + "criteria": { + "cool": "Turn on air conditioning", + "light": "Turn on the lights", + "wait": "Keep the current settings" + } + }, + "occupied": { + "type": "boolean", + "instructions": "Someone is home." + }, + "heat": { + "type": "score", + "instructions": "Classify how far the room temperature exceeds the target.", + "criteria": [ + "At or below the target", + "Above the target by at most 3 degrees", + "Above the target by more than 3 degrees" + ] + } + } + }] +} +``` + +Run all questions in one forward pass: + +```bash +CUDA_VISIBLE_DEVICES=0 python scripts/predict_toy_decisions.py \ + --checkpoint-dir path/to/downloaded/snapshot \ + --input request.json \ + --output predictions.json \ + --batch-questions 0 --precision bf16 --temperature 1 +``` + +Add more objects to `states` to evaluate multiple states together. The response contains the distributions, selected choices, expected scores, and execution counters. + +For a persistent HTTP service: + +```bash +CUDA_VISIBLE_DEVICES=0 python scripts/serve_decisions.py \ + --checkpoint-dir path/to/downloaded/snapshot \ + --web-root web --port 8765 --precision bf16 +``` + +Open `http://127.0.0.1:8765` or send requests to `POST /api/evaluate`. + +## Stage 1 initialization checkpoint + +The `stage1/` directory contains an initialization checkpoint for further training with the same model architecture. Download it separately: + +```python +from pathlib import Path +from huggingface_hub import snapshot_download + +snapshot_dir = snapshot_download( + repo_id="C-Tianyu/NanoJev", + allow_patterns=[ + "stage1/best.safetensors", + "stage1/config.json", + "stage1/tokenizer/*", + "stage1/backbone_config/*", + ], +) +stage1_dir = Path(snapshot_dir) / "stage1" +print(stage1_dir) +``` + +Pass this directory to `scripts/train_pipeline_decisions.py` with `--init-checkpoint path/to/stage1`. The trainer loads that checkpoint's architecture and parameters and starts a new optimizer. Use the root checkpoint for the earlier navigation benchmark; select a game variant for the new showcases. + +## Download the data + +```python +from huggingface_hub import snapshot_download + +data_dir = snapshot_download( + repo_id="C-Tianyu/NanoJev-Data", + repo_type="dataset", +) +print(data_dir) +``` + +The [pipeline guide](https://github.com/TianyuCodings/NanoJev/blob/main/research/pipeline_runbook.md) covers the JSONL target-distribution interface, data generation, training, evaluation, and serving. + +## Checkpoint details + +| Item | Value | +|---|---| +| Backbone | `Qwen/Qwen3-0.6B` | +| Backbone revision | `c1899de289a04d12100db370d81485cdf75e47ca` | +| Final weights | `best.safetensors` | +| Final SHA256 | `fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28` | +| Stage 1 weights | `stage1/best.safetensors` | +| Stage 1 SHA256 | `231b5178098477d9f82ae2cf38786d9ab8e0a80a019f009e55b772a66882cec4` | diff --git a/backbone_config/config.json b/backbone_config/config.json new file mode 100644 index 0000000000000000000000000000000000000000..b815be77563f3eef88baf4b4c8ae59706e618ad2 --- /dev/null +++ b/backbone_config/config.json @@ -0,0 +1,63 @@ +{ + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "bos_token_id": 151643, + "dtype": "float32", + "eos_token_id": 151645, + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 40960, + "max_window_layers": 28, + "model_type": "qwen3", + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "pad_token_id": null, + "rms_norm_eps": 1e-06, + "rope_parameters": { + "rope_theta": 1000000, + "rope_type": "default" + }, + "sliding_window": null, + "tie_word_embeddings": true, + "transformers_version": "5.17.0", + "use_cache": false, + "use_sliding_window": false, + "vocab_size": 151936 +} diff --git a/best.safetensors b/best.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..644b4508b0468795f284b78508ee721dc65a29b6 --- /dev/null +++ b/best.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fff62d1412685c1714eaa386acb603f9690371fb3cc8ad03dc41319302597c28 +size 2385039280 diff --git a/config.json b/config.json new file mode 100644 index 0000000000000000000000000000000000000000..f8be1ca8427b0d4791f1086b105e7c7f22826e86 --- /dev/null +++ b/config.json @@ -0,0 +1,66 @@ +{ + "input": "/home/rwang/openjev_codex_20260917/research/private_navigation_v3/views/coords_multi.jsonl", + "output_dir": "/home/rwang/openjev_codex_20260917/runs/v3_teacher_coords_multi_seed17", + "model": "Qwen/Qwen3-0.6B", + "revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "init_checkpoint": "/home/rwang/openjev_codex_20260917/runs/v2_teacher_seed17", + "objective": "teacher", + "set_head": "attention", + "steps": 1200, + "head_steps": 0, + "batch_questions": 12, + "microbatch_questions": 4, + "max_microbatch_tokens": 6000, + "eval_every": 50, + "max_length": 512, + "seed": 17, + "backbone_lr": 2e-05, + "head_lr": 0.0002, + "head_warmup_lr": 0.001, + "precision": "bf16", + "disable_native_triton": true, + "validate_only": false, + "self_check": false, + "schema_version": "openjev-decision-pipeline-v1", + "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "initialization": "local DecisionModel warm start, fresh optimizer", + "data_sha256": { + "/home/rwang/openjev_codex_20260917/research/private_navigation_v3/views/coords_multi.jsonl": "3a638666206e8ffacadd8beeaf0246687e546716deb8b06adf5e23a228890cdc" + }, + "deps": { + "torch": "2.14.0", + "transformers": "5.17.0", + "safetensors": "0.8.0" + }, + "gpu": "NVIDIA A100-SXM4-80GB", + "parameter_storage": "float32", + "forward_autocast": "bfloat16", + "parameter_count": 596250498, + "train_questions": 10105, + "all_train_questions": 10152, + "schema_counts": { + "records": 3984, + "questions_by_split": { + "train": 10152, + "dev": 360, + "calibration": 360, + "test": 720, + "ood": 360 + }, + "eligible_by_split_objective": { + "calibration/gold_distribution": 360, + "calibration/teacher": 359, + "dev/gold_distribution": 360, + "dev/teacher": 357, + "ood/gold_distribution": 360, + "ood/teacher": 357, + "test/gold_distribution": 720, + "test/teacher": 715, + "train/gold_distribution": 10152, + "train/teacher": 10105 + } + }, + "loss": "per-question forward soft CE, equal question weight; complete candidate normalization", + "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection", + "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing" +} diff --git a/stage1/backbone_config/config.json b/stage1/backbone_config/config.json new file mode 100644 index 0000000000000000000000000000000000000000..b815be77563f3eef88baf4b4c8ae59706e618ad2 --- /dev/null +++ b/stage1/backbone_config/config.json @@ -0,0 +1,63 @@ +{ + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "bos_token_id": 151643, + "dtype": "float32", + "eos_token_id": 151645, + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 40960, + "max_window_layers": 28, + "model_type": "qwen3", + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "pad_token_id": null, + "rms_norm_eps": 1e-06, + "rope_parameters": { + "rope_theta": 1000000, + "rope_type": "default" + }, + "sliding_window": null, + "tie_word_embeddings": true, + "transformers_version": "5.17.0", + "use_cache": false, + "use_sliding_window": false, + "vocab_size": 151936 +} diff --git a/stage1/best.safetensors b/stage1/best.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..df326108ac4fdaf5db10e9c80a0f238d35d54058 --- /dev/null +++ b/stage1/best.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:231b5178098477d9f82ae2cf38786d9ab8e0a80a019f009e55b772a66882cec4 +size 2385039280 diff --git a/stage1/config.json b/stage1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..b2931899c2ae7ba4203854683891e9466e88d4e6 --- /dev/null +++ b/stage1/config.json @@ -0,0 +1,66 @@ +{ + "input": "/home/rwang/openjev_codex_20260917/research/private_pipeline_v2/merged.jsonl", + "output_dir": "/home/rwang/openjev_codex_20260917/runs/v2_teacher_seed17", + "model": "Qwen/Qwen3-0.6B", + "revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "init_checkpoint": null, + "objective": "teacher", + "set_head": "attention", + "steps": 600, + "head_steps": 12, + "batch_questions": 12, + "microbatch_questions": 4, + "max_microbatch_tokens": 6000, + "eval_every": 50, + "max_length": 512, + "seed": 17, + "backbone_lr": 2e-05, + "head_lr": 0.0002, + "head_warmup_lr": 0.001, + "precision": "bf16", + "disable_native_triton": true, + "validate_only": false, + "self_check": false, + "schema_version": "openjev-decision-pipeline-v1", + "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "initialization": "pretrained backbone with new decision head", + "data_sha256": { + "/home/rwang/openjev_codex_20260917/research/private_pipeline_v2/merged.jsonl": "7294765b80e751fc5aee7aba906b28a8ea80d6f147253d3f6c3e4f491de0b2d7" + }, + "deps": { + "torch": "2.14.0", + "transformers": "5.17.0", + "safetensors": "0.8.0" + }, + "gpu": "NVIDIA A100-SXM4-80GB", + "parameter_storage": "float32", + "forward_autocast": "bfloat16", + "parameter_count": 596250498, + "train_questions": 4133, + "all_train_questions": 4152, + "schema_counts": { + "records": 2312, + "questions_by_split": { + "train": 4152, + "dev": 552, + "calibration": 552, + "test": 1104, + "ood": 576 + }, + "eligible_by_split_objective": { + "calibration/gold_distribution": 552, + "calibration/teacher": 550, + "dev/gold_distribution": 552, + "dev/teacher": 548, + "ood/gold_distribution": 576, + "ood/teacher": 573, + "test/gold_distribution": 1104, + "test/teacher": 1100, + "train/gold_distribution": 4152, + "train/teacher": 4133 + } + }, + "loss": "per-question forward soft CE, equal question weight; complete candidate normalization", + "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection", + "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing" +} diff --git a/stage1/tokenizer/chat_template.jinja b/stage1/tokenizer/chat_template.jinja new file mode 100644 index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4 --- /dev/null +++ b/stage1/tokenizer/chat_template.jinja @@ -0,0 +1,89 @@ +{%- if tools %} + {{- '<|im_start|>system\n' }} + {%- if messages[0].role == 'system' %} + {{- messages[0].content + '\n\n' }} + {%- endif %} + {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within XML tags:\n" }} + {%- for tool in tools %} + {{- "\n" }} + {{- tool | tojson }} + {%- endfor %} + {{- "\n\n\nFor each function call, return a json object with function name and arguments within XML tags:\n\n{\"name\": , \"arguments\": }\n<|im_end|>\n" }} +{%- else %} + {%- if messages[0].role == 'system' %} + {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }} + {%- endif %} +{%- endif %} +{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %} +{%- for message in messages[::-1] %} + {%- set index = (messages|length - 1) - loop.index0 %} + {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('') and message.content.endswith('')) %} + {%- set ns.multi_step_tool = false %} + {%- set ns.last_query_index = index %} + {%- endif %} +{%- endfor %} +{%- for message in messages %} + {%- if message.content is string %} + {%- set content = message.content %} + {%- else %} + {%- set content = '' %} + {%- endif %} + {%- if (message.role == "user") or (message.role == "system" and not loop.first) %} + {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }} + {%- elif message.role == "assistant" %} + {%- set reasoning_content = '' %} + {%- if message.reasoning_content is string %} + {%- set reasoning_content = message.reasoning_content %} + {%- else %} + {%- if '' in content %} + {%- set reasoning_content = content.split('')[0].rstrip('\n').split('')[-1].lstrip('\n') %} + {%- set content = content.split('')[-1].lstrip('\n') %} + {%- endif %} + {%- endif %} + {%- if loop.index0 > ns.last_query_index %} + {%- if loop.last or (not loop.last and reasoning_content) %} + {{- '<|im_start|>' + message.role + '\n\n' + reasoning_content.strip('\n') + '\n\n\n' + content.lstrip('\n') }} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- if message.tool_calls %} + {%- for tool_call in message.tool_calls %} + {%- if (loop.first and content) or (not loop.first) %} + {{- '\n' }} + {%- endif %} + {%- if tool_call.function %} + {%- set tool_call = tool_call.function %} + {%- endif %} + {{- '\n{"name": "' }} + {{- tool_call.name }} + {{- '", "arguments": ' }} + {%- if tool_call.arguments is string %} + {{- tool_call.arguments }} + {%- else %} + {{- tool_call.arguments | tojson }} + {%- endif %} + {{- '}\n' }} + {%- endfor %} + {%- endif %} + {{- '<|im_end|>\n' }} + {%- elif message.role == "tool" %} + {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %} + {{- '<|im_start|>user' }} + {%- endif %} + {{- '\n\n' }} + {{- content }} + {{- '\n' }} + {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %} + {{- '<|im_end|>\n' }} + {%- endif %} + {%- endif %} +{%- endfor %} +{%- if add_generation_prompt %} + {{- '<|im_start|>assistant\n' }} + {%- if enable_thinking is defined and enable_thinking is false %} + {{- '\n\n\n\n' }} + {%- endif %} +{%- endif %} \ No newline at end of file diff --git a/stage1/tokenizer/tokenizer.json b/stage1/tokenizer/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2 --- /dev/null +++ b/stage1/tokenizer/tokenizer.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506 +size 11422650 diff --git a/stage1/tokenizer/tokenizer_config.json b/stage1/tokenizer/tokenizer_config.json new file mode 100644 index 0000000000000000000000000000000000000000..33134953c69ef0d8f6d87a813851b965fdbe84bf --- /dev/null +++ b/stage1/tokenizer/tokenizer_config.json @@ -0,0 +1,30 @@ +{ + "add_prefix_space": false, + "backend": "tokenizers", + "bos_token": null, + "clean_up_tokenization_spaces": false, + "eos_token": "<|im_end|>", + "errors": "replace", + "extra_special_tokens": [ + "<|im_start|>", + "<|im_end|>", + "<|object_ref_start|>", + "<|object_ref_end|>", + "<|box_start|>", + "<|box_end|>", + "<|quad_start|>", + "<|quad_end|>", + "<|vision_start|>", + "<|vision_end|>", + "<|vision_pad|>", + "<|image_pad|>", + "<|video_pad|>" + ], + "is_local": false, + "local_files_only": true, + "model_max_length": 131072, + "pad_token": "<|endoftext|>", + "split_special_tokens": false, + "tokenizer_class": "Qwen2Tokenizer", + "unk_token": null +} diff --git a/tokenizer/chat_template.jinja b/tokenizer/chat_template.jinja new file mode 100644 index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4 --- /dev/null +++ b/tokenizer/chat_template.jinja @@ -0,0 +1,89 @@ +{%- if tools %} + {{- '<|im_start|>system\n' }} + {%- if messages[0].role == 'system' %} + {{- messages[0].content + '\n\n' }} + {%- endif %} + {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within XML tags:\n" }} + {%- for tool in tools %} + {{- "\n" }} + {{- tool | tojson }} + {%- endfor %} + {{- "\n\n\nFor each function call, return a json object with function name and arguments within XML tags:\n\n{\"name\": , \"arguments\": }\n<|im_end|>\n" }} +{%- else %} + {%- if messages[0].role == 'system' %} + {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }} + {%- endif %} +{%- endif %} +{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %} +{%- for message in messages[::-1] %} + {%- set index = (messages|length - 1) - loop.index0 %} + {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('') and message.content.endswith('')) %} + {%- set ns.multi_step_tool = false %} + {%- set ns.last_query_index = index %} + {%- endif %} +{%- endfor %} +{%- for message in messages %} + {%- if message.content is string %} + {%- set content = message.content %} + {%- else %} + {%- set content = '' %} + {%- endif %} + {%- if (message.role == "user") or (message.role == "system" and not loop.first) %} + {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }} + {%- elif message.role == "assistant" %} + {%- set reasoning_content = '' %} + {%- if message.reasoning_content is string %} + {%- set reasoning_content = message.reasoning_content %} + {%- else %} + {%- if '' in content %} + {%- set reasoning_content = content.split('')[0].rstrip('\n').split('')[-1].lstrip('\n') %} + {%- set content = content.split('')[-1].lstrip('\n') %} + {%- endif %} + {%- endif %} + {%- if loop.index0 > ns.last_query_index %} + {%- if loop.last or (not loop.last and reasoning_content) %} + {{- '<|im_start|>' + message.role + '\n\n' + reasoning_content.strip('\n') + '\n\n\n' + content.lstrip('\n') }} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- if message.tool_calls %} + {%- for tool_call in message.tool_calls %} + {%- if (loop.first and content) or (not loop.first) %} + {{- '\n' }} + {%- endif %} + {%- if tool_call.function %} + {%- set tool_call = tool_call.function %} + {%- endif %} + {{- '\n{"name": "' }} + {{- tool_call.name }} + {{- '", "arguments": ' }} + {%- if tool_call.arguments is string %} + {{- tool_call.arguments }} + {%- else %} + {{- tool_call.arguments | tojson }} + {%- endif %} + {{- '}\n' }} + {%- endfor %} + {%- endif %} + {{- '<|im_end|>\n' }} + {%- elif message.role == "tool" %} + {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %} + {{- '<|im_start|>user' }} + {%- endif %} + {{- '\n\n' }} + {{- content }} + {{- '\n' }} + {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %} + {{- '<|im_end|>\n' }} + {%- endif %} + {%- endif %} +{%- endfor %} +{%- if add_generation_prompt %} + {{- '<|im_start|>assistant\n' }} + {%- if enable_thinking is defined and enable_thinking is false %} + {{- '\n\n\n\n' }} + {%- endif %} +{%- endif %} \ No newline at end of file diff --git a/tokenizer/tokenizer.json b/tokenizer/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2 --- /dev/null +++ b/tokenizer/tokenizer.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506 +size 11422650 diff --git a/tokenizer/tokenizer_config.json b/tokenizer/tokenizer_config.json new file mode 100644 index 0000000000000000000000000000000000000000..9a87e3822403fb40a9524bb75edb66c6f0585750 --- /dev/null +++ b/tokenizer/tokenizer_config.json @@ -0,0 +1,30 @@ +{ + "add_prefix_space": false, + "backend": "tokenizers", + "bos_token": null, + "clean_up_tokenization_spaces": false, + "eos_token": "<|im_end|>", + "errors": "replace", + "extra_special_tokens": [ + "<|im_start|>", + "<|im_end|>", + "<|object_ref_start|>", + "<|object_ref_end|>", + "<|box_start|>", + "<|box_end|>", + "<|quad_start|>", + "<|quad_end|>", + "<|vision_start|>", + "<|vision_end|>", + "<|vision_pad|>", + "<|image_pad|>", + "<|video_pad|>" + ], + "is_local": true, + "local_files_only": true, + "model_max_length": 131072, + "pad_token": "<|endoftext|>", + "split_special_tokens": false, + "tokenizer_class": "Qwen2Tokenizer", + "unk_token": null +} diff --git a/variants/events_brier_seed17/backbone_config/config.json b/variants/events_brier_seed17/backbone_config/config.json new file mode 100644 index 0000000000000000000000000000000000000000..b815be77563f3eef88baf4b4c8ae59706e618ad2 --- /dev/null +++ b/variants/events_brier_seed17/backbone_config/config.json @@ -0,0 +1,63 @@ +{ + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "bos_token_id": 151643, + "dtype": "float32", + "eos_token_id": 151645, + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 40960, + "max_window_layers": 28, + "model_type": "qwen3", + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "pad_token_id": null, + "rms_norm_eps": 1e-06, + "rope_parameters": { + "rope_theta": 1000000, + "rope_type": "default" + }, + "sliding_window": null, + "tie_word_embeddings": true, + "transformers_version": "5.17.0", + "use_cache": false, + "use_sliding_window": false, + "vocab_size": 151936 +} diff --git a/variants/events_brier_seed17/best.safetensors b/variants/events_brier_seed17/best.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ce4269edd5714a3ac91ad7a53d247450a5955721 --- /dev/null +++ b/variants/events_brier_seed17/best.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a8d8b14c97743eed4310d461a7ae5cff0400cb87eaa6dc1f8c5dc3c59988651 +size 2385039280 diff --git a/variants/events_brier_seed17/config.json b/variants/events_brier_seed17/config.json new file mode 100644 index 0000000000000000000000000000000000000000..5c13af929737f4d2d7307853b2f116d98006a875 --- /dev/null +++ b/variants/events_brier_seed17/config.json @@ -0,0 +1,69 @@ +{ + "input": "data/scaled_games_v4/events", + "output_dir": "runs/events_brier_seed17", + "model": "Qwen/Qwen3-0.6B", + "revision": "main", + "init_checkpoint": "/home/rwang/openjev_codex_20260917/runs/v3_teacher_coords_multi_seed17", + "objective": "observed_outcome", + "loss": "brier", + "reward_samples": 32, + "gradient_checkpointing": true, + "set_head": "attention", + "steps": 100, + "head_steps": 0, + "batch_questions": 16, + "microbatch_questions": 4, + "max_microbatch_tokens": 16384, + "eval_every": 50, + "max_length": 8192, + "seed": 17, + "backbone_lr": 2e-05, + "head_lr": 0.0002, + "head_warmup_lr": 0.001, + "precision": "bf16", + "disable_native_triton": true, + "validate_only": false, + "self_check": false, + "schema_version": "openjev-decision-pipeline-v1", + "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "initialization": "local DecisionModel warm start, fresh optimizer", + "data_sha256": { + "data/scaled_games_v4/events/train.jsonl": "4cebf698e4b71dc37d4d0e985b6577dd565ac370ac82b6f47f375aa6af69f8c9", + "data/scaled_games_v4/events/dev.jsonl": "0c352ee1064a5fb5d2aacad59cbea57afb897886d3d09e5ca108c866a5f893db", + "data/scaled_games_v4/events/calibration.jsonl": "4a8a8a2b81c3c46df6bb4c8690e97374dd0165cbef40566bb7f87c1f425d5562", + "data/scaled_games_v4/events/test.jsonl": "884bec71891668199925809d8708347f56f7bf06e8363e4a436347967b1423e3", + "data/scaled_games_v4/events/ood.jsonl": "3c00d845524634ca3b34559761cdaf0c52b7b7633d5517d1e7ee7f31697cea48" + }, + "deps": { + "torch": "2.14.0", + "transformers": "5.17.0", + "safetensors": "0.8.0" + }, + "gpu": "NVIDIA A100-SXM4-80GB", + "parameter_storage": "float32", + "forward_autocast": "bfloat16", + "parameter_count": 596250498, + "train_questions": 1124, + "all_train_questions": 1124, + "schema_counts": { + "records": 2356, + "questions_by_split": { + "train": 1124, + "dev": 372, + "calibration": 368, + "test": 364, + "ood": 128 + }, + "eligible_by_split_objective": { + "calibration/gold_distribution": 368, + "dev/gold_distribution": 372, + "ood/gold_distribution": 128, + "test/gold_distribution": 364, + "train/gold_distribution": 1124 + } + }, + "loss_description": "per-question proper loss; complete candidate normalization", + "rlcd_candidate": null, + "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection", + "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing" +} diff --git a/variants/events_brier_seed17/summary.json b/variants/events_brier_seed17/summary.json new file mode 100644 index 0000000000000000000000000000000000000000..851244a8d64e7d333fab7071f81b3fd269ec950c --- /dev/null +++ b/variants/events_brier_seed17/summary.json @@ -0,0 +1,136 @@ +{ + "best_step": 100, + "best_dev_target_ce": 0.4421807345122086, + "selected_on": "dev target CE", + "objective": "observed_outcome", + "metrics_by_split": { + "dev": { + "questions": 372, + "objective": "observed_outcome", + "eligible_questions": 372, + "target_ce": 0.4421807345122086, + "target_kl": 0.4421807345122086, + "target_tv": 0.2684592128219396, + "by_target_kind": { + "observed_outcome": { + "n": 372, + "ce": 0.4421807345122086, + "kl": 0.4421807345122086, + "tv": 0.2684592128219396 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 372, + "ce": 0.46751975308439736, + "kl": 0.19751716047686577, + "tv": 0.17476332997830224 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "calibration": { + "questions": 368, + "objective": "observed_outcome", + "eligible_questions": 368, + "target_ce": 0.4847698677293702, + "target_kl": 0.4847698677293702, + "target_tv": 0.28490406827388554, + "by_target_kind": { + "observed_outcome": { + "n": 368, + "ce": 0.4847698677293702, + "kl": 0.4847698677293702, + "tv": 0.28490406827388554 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 368, + "ce": 0.4790668897887141, + "kl": 0.19277057070800374, + "tv": 0.17281581351777625 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "test": { + "questions": 364, + "objective": "observed_outcome", + "eligible_questions": 364, + "target_ce": 0.45052489631248066, + "target_kl": 0.45052489631248066, + "target_tv": 0.2634734042022083, + "by_target_kind": { + "observed_outcome": { + "n": 364, + "ce": 0.45052489631248066, + "kl": 0.45052489631248066, + "tv": 0.2634734042022083 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 364, + "ce": 0.46949283429897337, + "kl": 0.208722033972855, + "tv": 0.1715940341122606 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "ood": { + "questions": 128, + "objective": "observed_outcome", + "eligible_questions": 128, + "target_ce": 0.3166816048577138, + "target_kl": 0.3166816048577138, + "target_tv": 0.21465602967029523, + "by_target_kind": { + "observed_outcome": { + "n": 128, + "ce": 0.3166816048577138, + "kl": 0.3166816048577138, + "tv": 0.21465602967029523 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 128, + "ce": 0.3313783618564118, + "kl": 0.0978911402197844, + "tv": 0.11505404497567416 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + "temperature": 1.0, + "temperature_fitted": false, + "training_seconds": 94.21064267028123, + "max_gpu_allocated_gb": 11.964137984 +} diff --git a/variants/events_brier_seed17/target_audit.json b/variants/events_brier_seed17/target_audit.json new file mode 100644 index 0000000000000000000000000000000000000000..af673c3c3971ef76edb29b4a5af29ed643785342 --- /dev/null +++ b/variants/events_brier_seed17/target_audit.json @@ -0,0 +1,40054 @@ +[ + { + "id": "event_v4:4aecd3d8d87778bab6faae73:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a5cbe3a96ff9859f1afbdb0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1fbbe349381af7a541cceb35:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4db212f04284ad6caa8a4724:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c1e2ae6e3efa6052c93e30c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5378d5f62c640d6e0fe36b0b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f24e884a3b8605a927a55f83:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f458277aaf7786d8fe20847e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c644d0d083026cec6c153e9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ccbf6b8f9a588cc5089fa918:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93036e72f8c9f2d9ea0f0b02:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34792163cd9a8f97c1f4603b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69fddcedb520a75892669dbc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb610da24b02cd384933ddff:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdb45bd714f3c6fff339413f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6b7a7d7738d1c2a9874105d7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b083ef1931c7071e74652dc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fe27b46da3da7e169a4927d0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:196a1da34c2b563ecb152fe6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ec30ea0adfeee4fc39690d83:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9fa06a502ac89600ed238558:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35674713480dd730da6c2556:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b247183873660302da213d06:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91f994fed8c8d4008fafd666:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:46a6e10623052a887d1c2205:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c986fc04c6321670f2dcf9d7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c32b797ced7f503623b6407:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c8f73319f7da5359e93be37:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:168f30e6158735dcfa4fbe3c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e69848dda7a61a77ec326915:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9583dbb10b63349ac8fbd042:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e4ecc9a3d8526c6c4c36d32:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:778d5649938871adf7e3fe56:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01b9d89efa545728f8e9ff91:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a221ee0f166a7fe69dce2d60:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8683abff8c9fda846ef54b5b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9d8e40c9559f1335400bb1c6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e345dd2850eaaca29fad266b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c83318f0abed62232f75b3b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ce9228ec2e6d3dcecd19718:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3156d2a5861f363f4847de3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9bfe659d2661f885e63407fb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:19cd809568aa6c69380928dd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9e9b833eedec6d2516b563f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80ba1b17757572d541430961:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e84d57de9d328c6af74168a7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aab346f5a68e8b0d9cb8efc6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c9bfa2a376a324b78b76000:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bbc9efd11d627b4cf460c754:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5836727011760fc93a5ce716:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0afdffd76f8f8d912a2a9810:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f049876cd8d7bc1c1700c97:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25dc1b4da5ac8363f6d93624:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fd21575f55b2a921f658d041:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a95f0f6eab2d1da5dd6ecabe:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa7005c6ed1f9d7125195daf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6d76c5dbc7687767aabe29ce:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:840c07180c531dee4b614733:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:496c9ca3430d9a20c6f7c687:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e3afb50a1a36f19d4ae0b24:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:930568d553de51b6a5d759c6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8069fef34727ee2a1902a848:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5245ced54ca26c5fadec7646:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a4a10738fe569740dd0f7bd5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a6fb81b728f8dfa47a7ace4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93719e4384a5461e7a7219bc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12482d3af71eb59bebc3ea10:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ebf4a770693beb8e19ebae4c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a793ea6fc8b9e5d41771c5f9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:17a051caafaf5a77999f760a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f14ba7d8996c6f860ead6fc1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c79f246342ca3183f470e4c9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcf61590f5cf4749aa1280c4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7928a43d654582bcc499f60a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:510689d0c457bfd6b4ad0a50:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47177f54b2ff163c6def1b2c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ff7f129a6f7732d7e4795e4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a807079a25b56c67ac66ccb2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b0436e119d314b5bf86c248:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdfd7b16eb8dff830b5547f8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df1da3e45292136ec837892a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:758a0dbee8da7e152982df83:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8d56339aea5d5a95a5b665f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a88a6a2f1f0e68bc15a24c34:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0147d79e60563e071a1dd396:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:086cdaf3d5ab2ea94df598dd:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c718fdb80210fd11b08e329:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab041a7207d659c5fef22d4a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96deeefaa3365cda24ea483d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b74f223bff9f3d72331b031:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ff604244fb4de2bd94f9aca:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8300f14e76d85672a8c70669:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:666ffe69714f5e4b0d8a1dd1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1674014088efcac3fac914c2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e87a568ac2f2e78b210e821f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:254f7551bbfb139006266b6e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83cd93517670fd958ecae5ee:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:634d8cd2256448b12a2ebc37:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:081c18fbe82a3702fe712cb6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:860afa1ef370b2e6f70bf40b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:182efededded37823893473f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3d7e5dd2445849704958bf60:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3094e2d4d0cb34b5bf1d209d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:901067852a870e821b821b6f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d3a0a64ff3146b2c4360e4b4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e856c6d0068ced00cf2a375:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98e210f7e13494da7390809f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29f555194a95a09dfaa16eff:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c33b69ffd0e5d993162a45a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c37e3a826fefc4b83f1afc5d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7aff4ff3d50892441a8f2162:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc2d86938d5220b4ec91b2ef:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7b7c87b3eceb318ec6b95df5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a46694e1cf315b823b32b1eb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1424aa2a069f7812473505b8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4532e9616b169212296aed5e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ef476d69e35e007ec2ed8bb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b295a533b5c8c0ec6a94727:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a76c9ad5103bb6f3d9659029:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a3c6b25ad42df48f7a7978e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a05fcaa4caa3a3115129fb2e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4dc051c0a01fa2169c8b9244:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41b15a05073f28387831b691:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0dc938f084a2494a91529d71:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee54ccca139e1cf56c89a462:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a54e3adc5cf2a695bd83f471:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36aa1f5ef31c1178a822b653:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a6f3af0186a53a52c1bb58ef:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e6c7466ac0e8dd4278a1cb8e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea5ba6ec793a1e9bd609b765:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93cadcde1301139898ef49e7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d09ef3a5bff29862f750ced9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddac902fe22413b8b46805f8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d973b63f3df4fc07d52ed79:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cae384c4a93e87b360c116de:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6cca008fadc530233719175:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:538249383f3cd0a4e1dd2e1d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0edb8696013ec30415cd3deb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:554d0d960937239f75a5af2d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:52b8b97eb0c8779ef5da3933:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2cd0d4d1a69ce0bc5b4c649f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2898792139f17a9c4384afdb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d3e082aa226d2f5e24dfe45:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0214fa91d48fa1b6a1f894dc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b62b892a97c579a76b0f53d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5203bbe1d03bd01e2ca478b8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cba64bb97c7ab78c0fe29324:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb8fe05a7b0c62029f2e6ee2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d7193659271488065d5003a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba98d4315a4cbedd4e2a4a59:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d255c6ef457e46563b4b57ba:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6af87b747f689c9b4888711:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67d1f85b1befe76def8fee51:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7d9482c148cf9b95ebc1caa:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1d3310aa8547411a90a3bb7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e6b9dfd3c07d3ffa2205fd6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1827722f96e17db2bd6fcd84:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ea73ea02edda40bc54c044c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b4920c9fd2938836826284f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e027fbed3245e5b946c91da:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01f48f85232115f71836ae92:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01ee9a28d0f9469c007b5cf5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac5276e3d4831247884eb8a5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a95af135a3716d9969ef0e21:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b2888c81393ceb3d87874a4f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62263c53b73cf9c7e1c14975:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9569d605fdc66bb6c76a935:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85c6a22076b02b74573a4a22:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:616b96b76ad67d4c9deafeb5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:00d0e1e49c64591e5f0b158b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f709122ceeccdf38ad920fb0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85a36ad246c92d82ed439aa8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a8173dea9319f4acbdcd0f8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c93ed08d1f079a56d20cb73:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89b8fc1389543cc85a9282cb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e88d9a95e134302a29ec302:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd8e11eb57380f46a4533532:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3382186e120c20e414cec7ae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:af836278eb0dae03ebe1653f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d801e9df52ed90515626e65e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2afaba5e89d1bf4e970cdb9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e46b02ee09f14a9af0ee7bc0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22b1e3da175260251826164e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7dc426e15a7fa672e29109a4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:002b433b16ecc07c398aa3a0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c752e38ffab6e0a3117b779:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e1f79395cebd5fb8251a78b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b0f6359c5359f4b9d09a0cd0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1f199a87641482c04355483:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a9c472da2ec4b16b5e252ad6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30e895a88955e8e1741c3cce:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a4a29c08ad1ef7361d6066c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9299075c97ddb34754fbf99b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61046bb8aa61abef3bf5c069:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e812c569fbd2ece72b6ebe12:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a1ca87d4133b30b1b878c04:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c6877dfd63449d3097ef947:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4a07f731d746b4a32a673d45:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4a54319b85b667f85c1f87b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3a8aa14cf117a3c0edde5594:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca30abbc0d1a82277a7df57f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bace6e281ab7070db1cc530f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9bb6c22f14d1dbd762946aa2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7c4f642e1f7ed06decd0ea0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40bdb7a9c10dff12db44a3cb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b117b65e1f5ae573455bc7a9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a54baff6329b66e6db767bf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5533620bcffa09321a5d0f83:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0b73a17456822038153a599:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0d6c54e96ac000bf11cfbd0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d13966c3efe0950ebbb60c0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40ec2c07a3cab13452482734:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e2f8cd7ab90d301bb1b973c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85bb2b2655da536beaa4df6b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7d124aa6d2ffca932231b33:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4fb18e94b72edbadd4726e49:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:199a56e347100553975c9f2f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c36fdd7d4b562bc197282912:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8691b9f59d2f19d6fa25801:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:833b703763757af5e4b68b35:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:52720ad7a21f333c077b2110:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d932e48b4fddfa51788a6355:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:252d4d472631153cef78cc8e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0168631aceec340c766a850:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b80331706205a42292cf1af9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:248fb4f0ee7f5be24f49a69b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b389bc2b0a17e007ec63866a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41b2f786c19ea8833f6f57e0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54e3862145e26d2759ba5ad2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ad5a0c8f22b27c96b4b7a111:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7b9677016d81a9a7b746276:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e78bb8f32a3125554b5eba2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d01821e72b6a2a46cf93409:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e983594b520fb7548ca6a6e5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81ad6fb442bdffac10981404:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50bc7c88108132050d595467:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:221e3b56cd290837c7f03085:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b759eabb218117d5b4b1afed:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:549f1b7f97212fa1f78a3c5b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2493d89ac0d2715ed9bfda35:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30bf9428df242ea371274770:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d8322cb4fb052fd1ff83efbf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c4fe6f890052ee219851997:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:04724d4d373009c3c1df0eab:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a3cd9d92de85f633a5dec4ce:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d8481054b13cb02eb33a1e9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e0898e9b40f457a17d9026a7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b53d5d0ac873cde248198428:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b426a808a043cb0bd822ca3f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f70a11c264b317566cd767a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:43f61961b03ca186cd8212b0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2477ca30a8a97a0e4d25416e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d914d7d82fb5354aa46d8b00:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fc045fed1f8537f357cf453:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80243610db5da348b2a7def2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcc4e6fbc8c6ed9fdf00212c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:987b74ea0bf071b910f7e810:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87a1cfef0ca559c4f66887a4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c17fc3799d7c063b5c4bfb3f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4a30c805073de74309d662d5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a41ab14b6ab104574384653:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d01cc3d9771456d0369920b5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb9c8eeff8bbce90cd863a26:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb3d2173c7b2328ec6752939:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0652e8b64a4ed5ba0d647126:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf59db09948fe96a90130fdf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c3adceac18a6d6e8c481e1b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba9a4a75a947f371d1f5a481:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9cb3ba48fc6d5384eb585f7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:292d86322c0c34e7f5c7f30e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aeefccbec51a478f2223b626:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd71354748063f8d30e80b4e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fbaf2c0c62622cc533b401b1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:239ef25f5d857b47d242d6f9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98252467d7707f75aabce2a2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4092f7aa8f3083cfde4c58a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68eb85b4f7c251c9d3263f6f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7361c28dd204b7690a52e4dd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:236c20ffd93d26e8a865b73c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f3fcb67d58c6cf28a08190b4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b2add6850aeeb55fafb5066:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e439ad1b03f92a35d30381b8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:74db8b8082d0ec7061b0dcdb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e595c3863ba7c0335d80f775:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1476189f52ae003116af8f2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:104463a9e824c39ca1c78ded:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:679deb7673d4ab6a1ba5806b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fbc6594a3fbc5d7ddf80247f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2668a50afb39f45a5708eb79:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d578a3b877bda7b37b471446:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:938ea4a3f43edfe150282098:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f694ff937dff30da47693fd3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:990d3cba4db1fc53302fec38:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc42a7f86d3a9a6eb121adab:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bfee33f042bb50631e54845:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:789715bdf6ab3a2df5417bbf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02e0e060d800c08a7feb46a0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9432f491c7a3010f7c860661:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e691b22b64708e4cad8a4a6e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9350dc6743fbf0e2f2189507:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61a94f388e9f8f2dbf14d11c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ed16481b7895086a6ae0910:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f26146bb5865ce7081895d45:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25ceb291cd0bf06ba1da25b9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e1a74bd5e20eb6c116f78865:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3550014400bc31035676c567:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6369c20a70752ff7ccd6688e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5e6366bb9d2e3f6488841909:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d89fbc2a6938a5591c102d0e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:807bd7bb545f3363c43bc83a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f5c0e4ae9fa09e328c450e8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81dc5e532c6f1e076f70eb89:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7bd683df9b5f1f554ccac45d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e97dd3e41a15710dc9dea1f8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:214b3693143e52555cf9d3f8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3f1debc5bf500f6a54ed3afc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab0b47508e943da5fc0bc22c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6b16b1e88101ab6ed3cbebbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4274505c3a6b1ea2236ea6b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c10e2d63cf9593db22f07112:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7615413c0accc18c251c8e66:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5b296bf9fbf65a3f5eb2335:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12de4df791b5835d3510a9f2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:605cbf99ca463999d7cc4493:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0df1da747fa694bcaa207045:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cbd2aeeb33bfca228cdb7f9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddb4b15cdce7f68834fc04c9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:71044f235933988434f26106:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e1b42e7bfeaae483aded5f2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1932357263b412187a3089c2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8da659044d73f614b7ea22e9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41b49028c2c8f400c9467e36:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:882efce7e9d263e6984460d6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcfd6d606d9043e8ce82e88d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d55f1182f70611ed3875206a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96f2217a5354b46068eace99:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f7fac7cd2de3945d1deb891:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:002a53a5aceda2ac2ce52bde:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1fd73bed926baca20befaaa0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c5196a2731657dd74eb325b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e85d3f46b855dad1c5f1ddb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:255697241911fb3c271cef5c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4048f2e42e8eb20762ccdd6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:138733578bc678ff32537417:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bf24a4008d41f915e1fe7f9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ceca5ba33881a0df60c9fefd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:99ca534f1179e15e7d58b376:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d68e4967f66d33dbd27f5a0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb2146f6dbbb9fd75f85953f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ab35fea550164a919ea8020:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0701685f4b125ab98b01f43e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:197a8ff15e4efdacf1d9d8d4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:570a426861edde554174aa2d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7ec7d81347630c4503285b9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:797e6d858c34074a543dccce:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b995f52aa70685be6dad342:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e22e5f2f2f18e13030c01be8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4d777f612ecf72fc257f59e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98a69fe82363f91bc871aa9f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e15a945584411fa1b32245fc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0505d0c90f7e5a0ef9db8097:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d2fdf33219471e56a1ce28a1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9cf1ec307abe4bf0da3ec2f1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a50eea7365b428382fcc3f0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68ec8b82a6b56d72d79afd9f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 289, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68964972963ceaa0251fd85c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 289, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6e2d754f6e8a30dc996b714:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 289, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:79e1eb53a8ad8b73f07e96f2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 289, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba7a4d97d9c5cd5a91633d7d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c01989785a6ed6406aba33c4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:349a5dcd9f4878e6b189e285:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50b714be855f405803e90240:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87b2593405d421d6b7a28dc4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c199d672a69e2d7636cb026:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8280d12ab7784126b42b0681:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:721054860789951044116121:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cefcae3de562b00d42c08dc4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30daabfe807e170a81e6eaaf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8905f1c83713bbda13b0616a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:358db027e477b84d72e96a03:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0822effc0b54089eacbdff44:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0df480a355860888d0a5602b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6135141e7e7b1db72833bffa:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2ae4acb563ca7fea8469e8d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0669a64365763d5c2f696236:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93db94cf414a8f414c8d615f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8dc21220b05303caded38e7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3423add5646023034abf055:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed250b8345eb1d699b876343:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11af4b1978a1f0af8421b399:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61b84dd7e63312b7d348bcb7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b43548aca5107e43226b56d5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34ee612afad1e99c2eb710ae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:108910315ccec0a3e244a7e9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac8cd90feabb6eec95839d72:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb6d76511fe15121ee9832fd:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92ff30de0217490919c52f65:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06a99381abaa185a6a9fd049:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4a247b41be9e261216e342a0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a774bf3a92f5935173d3d2ad:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30bca1ff0db749828f0cdcb6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdea874c3af71af330e8dd19:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8d874354308484a8eae3f0c3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0dc3f71b2a4699ddefe5d14:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ce616f50c50ea8f50311521:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e57fe3476a705cca6726fb0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdfa2566ffdf094b245294ef:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc273089829f9de3e5112353:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:975ab57af5710e33e2b59472:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d7dffaf2c214fe692bde8fe:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:17e3d8f8e49028635017cd03:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9fd92e91c1dff41dc113b20e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:302d27a5ea6a934d5ac42b00:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0dc4544416ed815d7aebee50:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2243504d61f3609cb6e5a783:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69c5810b4917634567135eae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92ec2c46cf996dc7f34871a5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68e607d993a088bae18435ef:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:215140a4cda26a73503d3a83:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f854dfabf546b974d504566:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea3b5cc1ea8122b10f5f86b2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1350d478ab3b32c8b9f6b2d4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:550880c731ebb9d5c09ec24d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ea00ebe76fa3029ac6c7fb7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:476cc169236574a7988d1204:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a55e8124c6fa131a565d4b0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:303d0afd60f97e56aff794a0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:14435878a8d37bbdcd2dd5d6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a5dc91b79c9a2a823af6c019:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:60fdb9d681067a1a91f0b72e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39cdec65c8851ecd19231221:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1e923c9fcc64fece7829151:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d725ba1fdce0dc5b09a394ca:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef90e306cbfce9424175126d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f6754250ce60629b2509439:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ece0427d8d7392715d70ec5c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b2901b20d25de32fd1f255a4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9563b40907302e25d13773b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:57fac8334503978da5a90a40:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:347d698cf044ca9da0e619fc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:53e1ed1fe556a8e1ca5c2c4e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01e0e74ed27a6dee36461958:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c8fd4f874d575a1de9108bf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d740012301484b159c0a0dbc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6c7874bfad93f70a1ead4e6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:030b79e3eaa3583b3ab8d5b2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d6540ee67a566ec60d881d48:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c85ba39c8672f608029f5774:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee2ce711eb4c97ce04e46bab:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8f7a4879de75203775386680:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18d809d6ec9834b79dd8a863:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:64be282f7f43715069cf8f74:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3a7ecb6565f50af24257dd6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb0db769fea9db7eeb4fee99:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:710df1cd8f2b3a69905f05b8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f370e94562544c6ef17f2dbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a00e5d8f9b4a206c81833d8f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3356a702168cbcdc97cda7bc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12e939a19ea31dc542251c43:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c536e43ba623c319d126fff:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:14c6f58596354ecb87920614:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2103a5c5f4a45b95e0d6194b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4f9de9d504d227b5a994dcc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02e2cde5d8ed423d57bd5ac2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:908b506f0688ec79e57590e3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38994ff5042b36b0c1c55692:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b0f0c71ffd2edf8da980852c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:046c8c58d982cc8d5a0492c4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4e0166391c4975325e16209:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85a6940d1acda164739ea4a6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:56eb58aaba3e308c1417b500:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c1ce6f064ad5b498c644532:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d68e01be9b4c7ee47e12ea8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba040a21090fb09f79e69665:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9806fe49684837640b12d62c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c1ddd083656018f58c197ae:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5c1e390d768827716f5df84:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:79649dbab1f24cf15a3ec4b8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30230f7ea8af16bb21d90e88:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1831705374f85e0a6e15efd4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5dc9447d071b881704cb146a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:118905145afd251a26d29e6d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6e96af1995885c22eeabc35:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f478c7ffc40d4f4a0f8a2a61:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4fafa439945d3eeb7c614c61:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:796e4152d808101e800aedaf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3610a9e1816e92532cc72c29:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e1d84d5ed67202e645eb63f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89c2f4199a905249e1082801:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:55febbbe1c3cf90506d136ee:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bc9947dd9ce76fa7a77e0671:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba3a43cad6fd0c928b4ea445:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7356066daa7cc23f0e0badac:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b575f801c426fbe0ab34b1e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d7036fb11be505f0a9b3a417:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6923d5f85c7161e63d26817:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8991a01dac500c3e224d55f1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e2b4f6fe5f3a3c9d0f7880d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cb61a080e4ce926cf069537:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6193292593bc19879ab8d578:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c234b0c1046e975e8fc5e09:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef728f9a49ee671e899af858:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e24f353136e07d3be10cb3f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91f7cf44076c6002bbe8bf3d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83548f2cc37415c0eff38d44:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:109ea27a31551beee79f4482:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e504926c21cdc72076f9db5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2359ed68991dc93c9796f8b2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:72173ff70c74944a07b4ce73:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a7c9684b7720da082c99716:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bb0aec73c3ebeffe25c7e9e9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:110e7c0c1c03f20e82096244:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fff52480c3aa521aeb29bf51:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91f323f74c229b1247c3b9b9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d1f435cd245a901f0684e31:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:56f96a9f41e776cbef2774c1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29ff8183f5bb30345cab4390:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68ae6bf18b181373f8d6d4e1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0c3f4339abf8944e23e5215:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:717557dd783541f595805baf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee59c6ae58afff53499628df:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83da5cf3d1b1ddf8fa763b5b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5c92a5701eb6668bc1f8fa9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e469a62013803fdf94f1f59:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c2a82df1e5467aa7dc6660ba:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:26687c4ad73cc5184df26a15:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fdb332d5f8e4a0d8f459118:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0990419723e11fcba0affcd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca629ed704a66b1357bac779:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcdf7e375cf726310e726038:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5399697cd98f72753ff4184:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ecefb3605c03f41a999129e3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:845e97f19a29261cce2314ed:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29f234ad5aeef176f1e44927:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aea75938c86c8105d0feda0f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbec129f76b8c83c86381f95:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b348cadf48a3d03624ea0ba5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4035835d41d16c3a1fa8b1e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fece30cc4adc4871484d458e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fe00df43f4fc0eb779df708c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02aa23c6f283014c1abe9676:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0fe3e35ecdb9a7802759f3f9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:32d2a54b10ddd91944c573a0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a662e8b20507e285a191bdc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8ab283396653bec8d845a1bd:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c472a5c6051d07553cfec417:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47b61fd930b156045842a577:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa5b804e708259e0c04ba99a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:90c51dbb35adcc95e4ab55df:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ba46f215c4d34e0853b6ded:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b2be4263a3dc775e51b23c0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5427484a4262ff3c7fbf9f79:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5d3f0b661708700e8809d61:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ac884599d6dff56fa18f63b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e696748406acd478daad6de5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f7ff868b5c88c0cc528c43e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3514d35e6a4c8c67dabd6266:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ae1c08f689bd686984af0b7f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23fbaf66af2fb9a5c43ac00c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cba82a24afb9a6c6230917da:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:42d501bd5d2639615f1bb65d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1a5c14dc584cdc571cb6207:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8f4bffd9564951db12676035:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f1785f4091164d7d3300e26:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:420d69c51d42af2700fdb212:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b46c387df42006311cd59d82:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bac467952578b22f282d77b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac815c6135999e5968b9143a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bbae94c54a2adedee667f34:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e195100b5bf4cdf9f9dce4d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f58ae09b633370a388beca57:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23e073489564cbc08d4bacff:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a9fa261051646e4c425f6bb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c2ec9a05e0056357c4a8eedb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f29c436c6df6cad38deb3b9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:09019c96afc17c3bf63c15f9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f667b9249bd4af8ee7f52b5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba6d482786393b7dfe37c6cf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2be7856c2e3c093585329281:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1fa20e1404323924dd8b65a7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92a812459196dfd57fd7c593:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d2145a80946516e84732608:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3313bc5d19b525468ec7bfbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cec6d95e463244ba34935e60:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16ae671da86347a735e04d27:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15d0937b9eb7bb64031ec594:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:48a65de2cc08cfe9b53b76c3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb634b1bb36bad706a7084b7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:babc582a0929c2d7c07c1c17:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc30f2c0273cbc1bb0be8fdd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0bb4b76303c187ae6a16176e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e4375c60569dfb66fb42573:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:af85993cc9a07a6e132ae3c5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e328f9c46b379e4a9b37d08:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e7e58cbf8db050114101723:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:177265c717257ef23e3d69f1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4250b698d7b9c10b232469b7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9dc149656015f6baf4f46219:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a726bb28414368e00ea7c716:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa1457e17b5514610771b43c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06652b5bc00184010dcb89ae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ac6b2b50471d709b003454d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b2293aeb8e3eaaeb1d16d0b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b0ac2acbf820ac72b0cd6487:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb37e69ed7dcdc938d2d52a5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67d5b01b97a26c9573f4d52a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a04687efc794322704ed778f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66bcbb7e600cdcaf38db10d8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3593695e6fa4bf47bf7eb60c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54dedf522236b05f475bf13a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:792aa4e64f9a2e723059bc0c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0ddf3066fe7a07c046f2376:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a3ed1939a8edca581359f60a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:27327bad4c5f81287d5139cb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f8cbd20dcf4f9ddc5c51d646:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:607aefb0c7d22d7bbcf0ac33:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0fb010b04e5ab00b94f9c722:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cb5cdd9c1b9dd9fcaa8e0c1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8857c7cc49ca8c72df0dc226:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f2dec7dadfa71e1ef0479e66:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0acc8b69146025cd4934c70:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:57f0093b52ec2b56430a80f6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0593b127ed99d3616c1584ee:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4cb5f335687cd8b201f6d4a6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:97ebe26d7497d43e821b8e28:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9cee84ca5c088d1b37392d15:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c4d22aadb729cb8b826be92:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd787b306f5b5f04cab14c66:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:94688c378413f969e8701a8f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8dc99806488ab6384e7aabce:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a3bb296ed9c28ff8b9a5d743:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2af1038af5283ce6600f3d2a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5738c688f78baf9bc6dddcde:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8f9796158617bd8c568a93d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4355fc17714a16342ec2b97:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9af375fccf968cba465ad0f8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:86dee5796e3794236a4ab2fa:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4cb9a623c1921ae41eb6e0d0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6b8c3864a7d66bf960feb6f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96e2c9083ebd21ea07e08c1c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b1ec99a0da91ebd52a4bc72:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:97d1a8e484488d2a6ed47ef9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c73dfff2e86789c2afb78b9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5747d09e72c2bdd2c192de3e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b219c24196b6b47e8990a71:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ce736a0e9a327a1e3ab28171:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df3045a4d9eb07f99fca7e99:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75c32f6c1b3f14c4bdca3a67:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0090f026f9e8985f16cef679:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:874fffbc6e8b724a715a1443:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7b8ec24d53082094ba6cfb7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2cb5e70aaa43cffa82c7e399:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c17e9d4a789e6976eb1205ed:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:17a10837f8c166221f9234fe:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f7df1957ce58aa1af8ec6295:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66f14148c22c4c918d32a583:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:068bcb7a7d9b80b0211ea29d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4567f0f02ac21971af6c2067:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5970165308c46d21f689248:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b820c12089d45b425a2f143f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8db344327e8832b431640f29:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30e5aacf3967d75bce836157:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2aacbffc4db777764d114fae:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9791481a2c8586a8b9359f89:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:150e2cd536a5bc1059f65dc1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:71034de81675fc16e59ca900:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:432c9d646994c0bc4a45c30a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9fc0ffc853df57b1434b3f93:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1cec312461e179bfe87fc6e9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc228262685c9af863e04f40:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0bca2d66187e300d825b0e2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:185cc5fb91e6085cf42bf0fc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ebbf8f64e89c61e0f31df873:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:256b930168cae56bd6137a15:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7ab006da8bd3d6f8f094feb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1dc1fb1bb2ab20d549e1b067:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:801cf4e3380a9253f7340871:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c8179a22901be601d02a0bfb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d22c006c056739961a9fcc4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ba08d18374086eb044a9a59:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1ae888b75b45fc321ae31b8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e182b61b80cdcd82a8afbcf3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f61c18f607c68048efc76767:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:760691b4a6b553e8e6753666:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5bc6da415ac3e3139c77604f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4f74c2cd88ba1556d014cc3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b362adce01db8b4de0df49f2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d91ef6762a4834e41e226d7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:da4878ece31bb24a97167d65:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f90237ebe0a2a0b99bda2998:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cbed873db6aab5eeace3694c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d56ef7f829097942226c38d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d56a2f30696951e16e25a68:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22d3b43e71ab03e8c809e636:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ce6114047f6b01f950d2c7c8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c68b5115e21bbffda6349ba:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d75093dc65ef08846569126a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:af9403a01d459574d8e44d99:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:204c301980a3c082bedfe1b4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96bedbaf1dc4a9c585bcfdec:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8cea37a22ec1d8811e1316ce:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e234ee54eaf6e14a7a3f023:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf8cac29d1b7416ff9a898b0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2fa298c5736a119c0f39ff7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4df98939ec82fbabd65766f7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30c9c386bd32f1dad5ba2fbe:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18eb3dba14ceb391c588078c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fae9510eab9a0c6118233776:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a45132adb7911e7510ad744:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c0f458485c6a396c24f942f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:20b549d9a6af728fe092ca68:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f9b6fa2af2d327776b882ae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b057646709e1db1debfd35d8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69cc59bd8e24338a96430932:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2307ad191e5cf6f9c229a277:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a08b246a5ddaddabb4b86f4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:799ded37832184f9b0ee3168:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcb0dd55df41f2ba318a1bb3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b952616a0a1f108d33bd9eb7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23e2df6c0d3866801ecb2c45:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3dcb436a16a6492943a2fe42:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f1af8e03283abc44dee7859:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:046c78af4f31b3e1ab065ce7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:708587746543a9710eac6c48:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d22aa8347ce5fb1a6dfabfd4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a7fa96d2e7cadf4a5f3b6b5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6637b24f2a43ddb345c43f71:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80890ddbc902f2364a603ec2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:60fcb58976b841328ff5a9bf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:229a9b464754862682004324:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddf5a73e9542bb34db9abbf3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6278dbc5ca5e99ab00b0144:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a300c51301010ff18ce3cb0f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bb87e2d44e695ba430954bbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5b9410e7bc5fcfd47057e921:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:185eb6d6fef5f701433b1220:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2dacec95a2422de285a8dbe7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3d31488468fcf83a2c803e4c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:76cd3dadfb8a076af5a66d49:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5ba5569d43760dc324e2833a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b714492a3f71948f1882a11:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92cf5eaf90736c9e783955ff:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62e0f1370c4c1b99cdc6e5fb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:26b949ebe9b83c3aa170003e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66d868a7bc10477729ef2a16:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 595, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:618fae9872bb340930b47b13:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 595, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c7cf7b900b8f387ab2ba4ca:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 595, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7566f28d2f41ae8fda8908da:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 595, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:049d0bc4ffad5871cc6c21e5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e85148c1ca8064cb491e1ac:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed45a43e07ea52f3474f3c6d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa72e2e3de57338165467204:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:496aebb810c5df7c3fb73488:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e32308b867c8e6087e133433:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e984765547431adc0fe178cc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e4274c9d2832ec90c5d381a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3389c4d6994e16bac4135b4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 597, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef0fb44a5c127f5022120f02:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 597, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1bd52436398b2ea1bdeb6a0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 597, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b30fd38a909a162fa7d896bb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 597, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0dc053390957a7149eb3d08:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3b41c608f1490ca652273ec:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8dbd0e73903b75845c88813c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddc48f37fc6e4f36738bc86e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c9592fcc0e2bd4b29222061:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a57cff3bef65fd2e9fbf58c3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c2f66b1dde9743bef20b0266:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:337b5e5fdcc4fab7d3c53ae5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3f605aed9542453a8aa3b21a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b83faf6baf4766897e6521ee:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df78fdf4f7e75de43738ad2e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f166af9dfb76b10ebf7f879b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a78ff1ab2f9b43ecfc7ff3b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ce7589b9818b5720cf2933c8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83345349bb580680ce39a90f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5c5d234c0bf971c21b140ac:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab4e71d67dba088cb00b6bc8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 545, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:200a325d944d71b733e57056:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 545, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a7a42b1c7fec28886a15d87:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 545, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9e494705f31d7673ac4f282:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 545, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:739fe045f64a353251f53255:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb79fae11d1e445c1a56ef56:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d7fbc6d82cc3b6338a9425a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1271717bd6c4ad47ca5ab540:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd275452a7b4561d073c7d2f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85b9ff553f1ee51eee4b2ee0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:95e6d29eea890206b64a2b9e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a141648fd4d5e30c21cfe742:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6bc574010c1c62fdb5c06ab5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0bf41df877960a294bd2dc3b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:988af17e89f3dbd0c2a6d8c9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:996dc372ff31a1dbe5f80051:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:091c3f8e4c36f93849ac9785:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:079f9ee4be2ddb457026a050:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8bb5f7b4bbe51dc82621e3c4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:17b0baa120987bb826e6293d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0299d1ecf01b99211039b2c8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3a5c95f52cb7249931ae8054:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df8c537529a9f98e6d25c493:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d4eae14009ac454c6dd3050:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:062eac1ba14e29549c90605f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:782a9ce4f823249e70efd6c2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b1501a6246bc51a2aa4dd4d7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d264e421eb3650ec6c88f72:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d292e83ed35deb88990c23a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b51485b0f6cecff843c3a3cf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a64958ccbaac405f2649562:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2bbca54f9fa3526037a45d5f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be9545bc7bad816745095a84:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d971b583c5f6549dda4a104b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0c431c169d5350813868fcc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c463c601d22ed6f862262e2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d38c8882612470ce3b7dab3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1017887fed6b638c2029360a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:669b7a1ee069c220b8bdb64a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:73c1d9c12699beff2c57db5c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:604f087db4066c54036e358b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:518fe21c787ca08f888b605b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eaf3895d8703c0acec712ad9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba8ce6c41c9ecde628f8f085:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:060e5d024b462f4ae3b4f91f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 571, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf7e6ec22c3105e56dc1b43d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 571, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:97f6f5a77d290fa5591db5e1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 571, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c211443d389ac0b2eaf2e0f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 571, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d617576e2f876effec9bb0a2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca9adaf9abb1b02a390ba105:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b855b583ca6b1e9350d9aa73:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18833593ad322bc2dcbea547:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a969259678ed90e865c28fbe:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b4535265f0a2e35d06e4f570:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb48a578a464a1251337a945:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50948d9010b0cf9be4705043:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c6993e692c8ea421debb519:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c71a6020390573b60dc5648:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34e76f4638ada5d0d6ac7788:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cfb6b092cc0c84712b1aaf61:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3fd046e082b53d5371704ac:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:edd3c2b78bd679cff92b51ef:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e23870d7b4eda111ef3a4af:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:73b4b18f273a7d95d9e765ad:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c79eb4114c40e995f3c39e21:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65a2de6b84eca499474c8690:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:42b1e81510ad3aa3c14b7cff:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36c0a1b384208e83ada34096:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:977b3d8d36ad484117f35b86:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fc20dd3298a3c6bd8fcaf067:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3dc84da1587d06906545f61c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7cb0ac616386bc05c529e74d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8848bebe5d1b3ba8cfd9e0cc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6cc84949cb57ffcc2e00fdae:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8aea77496507afce37ecb5a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf5c447bb2d53e2bc6f8be9b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c405410739966da6329ca882:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 574, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4590b4784259a94ec3af0811:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 574, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a1837d609911a827f872b9e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 574, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea3dfdae3aef291e7700e753:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 574, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc3ae0450bc3fc8e68368bd1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b396bfa34a9d31d23deefa3d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7cbbd588f069c8555e06a379:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3237771955dd6b30d41e8c39:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6df4b63890e6a372f2df7e9d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c152aa8999fa73ff99ef885:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b025645c9360bf3a899387f5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f34cc937b036b479071a9d0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5acbbde21967398410a1ad3a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb9dc3ab77485e3a990e3d56:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bc767500ce31f92bebd87dc4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1616008ece3befbcf7ac3456:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:984309771153005679f20451:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba71bfed0a26b0af1bdd2695:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ccedacf25cac3995814020c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b126e98cf18d82316644705:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01718a17f99bed558550af79:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0fd7b0b208e08b9f3344eb52:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22a3b10970c4237e0c8ed51d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c5a6ad773b9dac3e0651374:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2d169c2d770eafd56fd4758:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 470, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3a59d88fc557dc149634d73:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 470, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdf26d139fbebaaa044c2fa5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 470, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:055c56e08b4932985b34a473:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 470, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96571654b9ff81c52be288fa:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 471, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7196e90067afe6dd404174a6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 471, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:13b7758b6243cac5f4ce4f33:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 471, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29e879f91e5f944ae6fb86e9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 471, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:72d4a5c83ba0dc1ad7b77ffd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa1fa26103d5faf7c72e5608:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:074d4493dacd1dafc762a069:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e63620ad96cb32f54bf0a456:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b139a1805c6830a6e0346a13:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7028e58442b3c1ca5f7a2bab:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b48536a855b27c38c2b91f7b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1f1a3eb1f7ddbbdc865bcaf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bb7af29c3d9edb7c43eb5892:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4a4d7adebf309811c1e26a6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5d627257bd76f3bb531e065:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c8678893c0a76919aa83b1d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07501ccb8a07f49088214544:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e464a3a08ac314196aac74aa:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0027d23e3c9a1948f746974:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85b5063a4d5883992e6d8acb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f10fad0488934eb0b7414f32:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a56230a85cabeb26d046937d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7ea98cdff97172aa7909b3e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c6dbf140264ea67b36f0e22:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f253fa8955339167f5ef9e7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c63bbee2ca96de372c41de0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fbd1a1bae32dde189574541:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1be4d49642934fe08545764f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67f61a4add1339e5e069b739:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c346f8cacbacadb164856d6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bbcf2d453d8ff3bed96e9911:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e656e8fea427f33118c16387:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3f85c9e8efc59b4fcbbabcad:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ab01cf624903a190a6f9c45:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f5f538db7a194b4a2dd3e12:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6db135cb082032ee533b729d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b572035415c96f4ee0b9204d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7fd2cc20c2d4c0b6d47f2e1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0b994dc9230cf4cda2cb02e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50db9e62d29594e9b708142d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcf8541ce0550e598a5ba3dc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab8ae672e41046c1e9950315:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61a57e79294863efbaf08741:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e56b912c71d445003bfb4a5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:483889dfdd43b2b4b67d1f0b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:88c6ad09b0de5de5a96dc1a3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:618baa94527b6a8d7be08d3a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c84d1ea32238fa2cf7e9f3cf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d44348b24ab9996c8fa75bd8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85f2721c164d94b1de3aae76:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c3a4ccc6de6097845cb1a04:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a5ebb92e1491f02dd8513c5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b8dbbc2664243f529d57120:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5c96432e55ee99de9fb2b51:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:88f5725c37f73c24b68212cc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c26656968a0f61cf1c796605:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33faabbbfc6433ff4d689a22:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8bed7ccc4601dbf65880a3e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5cfb3ca0b69efdd6ffa3a4b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07307ec2d55120b44f4195f0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:287b4c2be4ce7f890ec460d4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d19e27c3ab06efbaca54ed8b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c21c6ed160ba9b2d5d46b2d9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4987475bb3c5a76a9cbecaf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ff71f7ca6922896cf864dbe:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:545469addf9c3e20dc5ed800:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3da02d11ab3427b378923240:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12097124a965cffcf6e69a6e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b4c70a5822dcde19ef83d72:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0fcd4787a5d6c923bccfd47c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4dbf1a5fdadeb368ee2d6fdf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cffe2b5d1d8fae8977315f86:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:615fa4d6aab97aee94c4e8fc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:147608a594478104c0d680e1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:45a81ded48065fad628676f3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8a1329ffe89c64471af3ff8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8fd926445bb1a824d991482c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ae74680c924ec9e11b897dec:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a538341eb215abf1380900a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc6390a98c2cf33ca8901588:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:341e910793242946a516b5d3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a451b75cbf44c73bf74336d4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab67bad44e9e5ecba285674c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1340a3c59769cdc31710736:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:600c4ef94cea891dc9f7fdb1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:959379badbee938823522893:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a958830f5b4e68527d62976:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f755b9e0473b052c0a3aa602:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b31a1d5a3ca6c179e0a8838:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dfe36dbe4acef0776455fc99:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40b79952844d2f3c619c1345:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fabe6d0c3e9a7212b5f81915:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f7598b99e691e09e3a1c0b0b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:94c7291587c7fdf8b9cd1f7f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d66f45b2ff8b346064a296e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e3d42be8eba9d63361c893f9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5ad7bbbe7642013833ec7e1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5908e06d34bf003d7320046f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02a91adfcaedc94cd3fc583a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01ad27da1a36c9cfcfeee4ed:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75e80b41dac1e7e7e6338d60:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d89f7f3845a780149a83083:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6fd62684837c029a58c01961:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40070ccc7f7ea6b00c53822b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a118cd61fb57f5213613afac:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47f135bd19c3e53e991d8ae1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89a6d774cfc7356312c86566:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c689bc24e86a52674e7a022d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7fa60af1f417869daa465c4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07f166996e83a5c86e7d0708:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c721d20f39e8a865b3e38dbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a17f849dd131552ee8b1c0f7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c903d96368e8adb351b9fad:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f3b1fd80f7ecdb5a513b560:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7bc3aafe23a827ce7d229340:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c191e4c033dc516e2def8b7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f9b23c73cb7510a5a2adc3c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3fad815b9daffbe49eecf24f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:48a166fed1e0248e0dddab44:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:63f6cac5125a6e46b2e48347:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91e20cf6a48504b4ac254a01:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:519b1d4725a45bb41806f62e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c21cc954aaf7abf7c59f3b7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ea4788f85caaadb0bead233:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:493ec62203276581a79afaa6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9837b7c810380c92f975e234:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3d525bb886f99fa5d17bfbac:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25d4490a9bce7ef88aa34685:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34db7cac091b269184b7468b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:24ec67ac5fd9e41a406e36bf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7c85a52d4d1ee808facc872:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:32f8d457f532d2dfa2906b7b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:03c379704bf9c1ebaef44f82:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 468, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:781e96970e5b9263f95e9471:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 468, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5cf07cea038f1538c06dfbc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 468, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b23126dbe3242eefd1712c3d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 468, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3212d633a8ddf2624acaa255:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3ccd8cc0a0446a25bd29e026:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c642da1c522c16fb79934626:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6780f90bd4ef0ab4c964c852:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7ce50cc4380b8a55d34f2ad:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4b222779e4ff9d85881cab3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2328e1873f3677593b45fe70:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3f8d269946cdd77250c0ca6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5bdf970cd0ba998b778b5677:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e33ab212ca290dda2901f5f1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a45b2db533539d94499ac60e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f50ed6b0f102353512c2c0f8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8e59a6e593d8418cb6710d4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22af5bbcd4d95a783b6dcc39:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:145fbd6ede95ce99e4bf7b3b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e2e801e11135fd6f8613f59e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c3802c59affcaaade9387bd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:945264ffb4ce0fa9ba9f62d9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:13ceb9c9e5fed7cd34292306:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c162b5b08fb13c438da4fcb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38df9ce8f0d3ae77cd5b3d34:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 557, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3d3d83e371282b03a9d672e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 557, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3387f9c50b8eb2eb5bd70c80:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 557, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3b888f524c2b32f762e2701:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 557, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b20e8aff695f7d24e9be2b28:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9bcb34864cc835f2b791e60a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d028051d64f42217987da93c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87ea8523e8c98555a7446120:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:99abb8844b2fc5468bddb7d7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87c347f39283b0d43c77b36b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:485e1836d263485f014bf040:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f8688508369fe71529e15548:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:462244b37ace954a228fb0e6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75e7bc080a392a087b2314f0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b43153d096085d3ad116033:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1aa4fde82e480d4ac83dc4c2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0da2d71ea98b488d4825c359:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a054761dccb525a4701407c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:591f3eb85d17de8b1ac7f19d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ecaaa213a8ac9a46b6270ad3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cbfe4b454560682d4dac3674:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3d0301825674c044655bf054:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31d62825e4e3faabd7cc068d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e981d07ed9b3bbbf0981a372:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b1ab97aefb5a673e3dc803e6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36772881471abb9fd040e64f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ad44eec89c2a9eb1f3f8eab4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7fdee6f4441672fc1c465c9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:43364de707aa5df6deee5816:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1999401b40125ac07cad81b7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca0dcd12281927da367a76f8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dda5d42b55829593d7fe9b0e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:727939aa5a329baf68d7a0e0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f95b5688118480526695ea5e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc66325c50583ee4fdc23b20:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:59e119ff87abe2b42127574f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6b24c14151cba9f11c6e787:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:94a76bed42f3b0e80089d539:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f9fa00b0697ea252e9a99625:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd40fb140166100c77265ef3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:74a3b1206e2a1eab5a52bf1f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cbaf99e91292389c5923b2fb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54a6e2e10c2f2cfc6aa6b776:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:afd58a1a0b150557c4f358b9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61fe10a39e7ab28f99a50e4a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f52f2584a7c72fc9442beaf3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d914395c0130d80e25f5b1e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9adc6a9552d422d6d88b93e7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22b5003ece8ebd2aa85dfb18:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:133375e6d8b9d78fde7fe0a9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f55652b7259c7ee47df3119:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:163467daec81d95ca3d8af23:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ad54d23d660eb5e629f7f7c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0af79b5ff2ab389db6a9b35a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8388537ee2384b30c2c6feb1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdc7364b4884b8836a3504d6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:833f6de9d608f30dab171f6e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c7c2daa5474d5a1a8bf11fb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:905af40eb3eb9c4955566266:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f18fb2abcbc23e0b64f52044:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61e1497ee9e585db875e4537:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:72db2db4d3ecbb56783e14cc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7de5151d93cb1052dc0fa91c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9aafc04b8b07637d4e6752b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c04098380a08cf3c5fd7dc4b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:839128e9da684bba271045b8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0cb40dc7be962f868cd0d15:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ff3ed1c28428f72ce3d54969:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30fd05c79c32add90cccda8a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a80205837474bb0fd10f296:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9dec1ed34053d646cd07921:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa50726b01107030094d50a3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:885ad0ef88ab51749e043fb1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f33b378decbed1eafe240310:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8efc83a4ba1f7cd1b0a3d137:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5324a7ccbb9088a70298bab3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:21bc249588c2781841c7149f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5960e3b44b02676823bc81b3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f886e82fd92bf0b8349c0131:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d595b4af79d59ee63a488656:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2908101dac2781f123f0ef68:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:628055a68b19320e37b5b9e1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ef5a13f24bb691a08eb4894:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef5aa99df625925ff164df73:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2fb5cc2e2e671c63ecea93cd:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c75bd0440912fbec035278e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c6963d8eeffcf39a12cdaf8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:42c14eb9a1972ccc06eeb024:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f398ce00bd97fc5300437285:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c95a318ca2c09aab72f0b60c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7af16f45221c1b25d43fdc5:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b34dbf346315bea049b3fb3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96b9aa6cd719032a9c18257c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4ec420319578ef4ec44166b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50090cc679034c7daa5d5dff:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aeb3a33c603129f6ad48990c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4003b845d1dba349c5b5de1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3cf63a5a4dc396ebd6a1e86d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5dee3df69aa1c8dc59132a45:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:248b93ba85c9cbfb12c43dd7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33f56e9af08bb8c54919f196:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5de3d14d983a2614a765b5bd:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:46e83cc9f0b107b1d4b5dba2:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d8237b555d94e916dbdc00b0:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5663e783bb398ea72082df3:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c223dfab341eb69c79616729:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:46404f67802b9d293994222a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3fac0b725736aa037a284eab:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16e5d6ce016de554cdbd90a9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf1d97efbf2d8584141f71ed:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8951b6f604ec215762575314:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d052df36863e4df125f517e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ee2cd16ec08fe398bfa2a5f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c2385f928b4ed908721b40af:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4e3db884b0dfade44dd9330:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8409b7e8224ab1eb841b57b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e31cc834f61ecd9cb5c3e740:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f1dd8cb4fe6ab50fd4b3e11:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c02af18406fff8411250e95:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0979efce03ce9419d8994a1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed40ee9cca77a8be8b9fbbab:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5307030384d5fe47ab60330a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:707535ae8ed979160e6b04c3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb4c68dac851f72cb8fabc88:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0952ec78fc2d73f0df518fa:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9291fe8ba094d83b4573830e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c87e0c0516ac7f4fa71e77f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ff4cc1040f6229f89180a4a4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6b873ee7e0f5fdb1abf77ee:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41c2f6007e0e8a0bdb0e0f71:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3146e1cce7399b8e28959db7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58c6d4dd95cab700b72b193c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:576db2fc3e9ac93770de6483:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f1b104a5bf261b57ee60430:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d761557c925635868a3fc76b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3191d9616353be78e31fcb2e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bd78639e23fd2263538d1c0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b9d7b74fb4d78591c2e65eb:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b9b1854124a8236ba783096:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54d3d040008e83ee425b1f6f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5a17f6e42901b9f1d1d3c8f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:224571a455d751118a1592a0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f0acbc9db544f0da7b03b9e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81e6efa00f081652c8b096a4:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96c11e9b0f4d5aef92569942:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:21935d79c6eb68227a4ad91a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f6937600ff421388af21a2e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:534fb0d488b7d85d398fbfa9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f0211ce19200b833f04fb7b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:310e0456c80d12247bf7d61f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4be135895b1292470b5613c7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:591c3cc1abd7e0baf57b6f56:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2dcbcfa3e1bf17697a7dc728:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5e23df8b5300f02cf83baa0b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df42f6342483435ef082cc08:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b581c4b5dd4a47ca358c58d0:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:00ad4502930c218eeaf66300:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:24598b2e1692167beb453b93:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6410bcc195012defc1780a3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f385b6f4f2ed1bb9687b6a26:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca558d1efedc7bc4c6d76a4c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5ac8020fe947017cb8dcc098:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:328a2a0aa9585e8b9d115453:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca5706b1e59c369c22972456:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:552728dca0751a5d929d866d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11f668887cb34681226e4eaf:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0276d049b418865881211dcd:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a5b02ca0a3827dd7e41cd7e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16ebb97c6418bf91877ddf82:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5a783eb6199ad75c4620232:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:429f555c250f7077556c6947:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bcd61b8ae841a07d4910f7f5:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e15fd669208dc30921c03369:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5455096fcc40f6e589bff2e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dd2ee3f2acfb76ab2546aa25:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ed526d0bc4f86b611cdb278:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f837f5e790c0ce258b9b477:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1bae6d5dfa00dea4a46ba23:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2538ecd39313177c5433cfad:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8423b15e5bd1044f29639afd:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e99b6575f88dec8d6c09f7d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ab549f45e7c611807b0c342:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31d8e8482febb32e62cf53a7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:49d20519fe9b492fd911b0eb:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6b15142fd829ac7e4583a9f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e768d2f79a8fe07650bdab4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d15a3dd3243ec9d39477ddf4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:de52fc939058cb7b8278b22b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6caf1df3ade6479c87147788:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:09541d962149d706515b7aa0:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16121dc0ed0ce041dc78f926:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:973d0d65ad51559ea03921b8:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87af4db526b7f881a4d5a315:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9d7a93f9a57dc7c27dd75e2e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:71c2e23c877efb17dc724b21:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01cbe0c184b8d989bdd66b8e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fd0c2983daf59eb4fb0300af:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fac77920146c5c8432abd7dd:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b54516c2ee133fb35b522bcb:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:675a9320f80d4b2f9fa53686:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:42a37ea561c0be6bc3533125:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89e0ba188ef8e6704a6d64dc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9cfdc0ced617220922eb71f6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39bd5569447e539c9e74aae2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a3df479a9e5186da2dddf934:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:19fa6556e3e5ffee61730514:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9d9a290f5daec2408097133:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8fe29088a567384e2e6bbea5:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5ea752773b4597b216c4798:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:952632a214ace986f3876d88:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:53d851e214605104686da4ae:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:436c2971fece8e8a1cd07b6c:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0820228a7229f63e299934c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:477d8ac01585dc486465ff0a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:306db963fb6d2627de3b7867:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7909d1e5faac480e7837c3f5:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fc59f5f020c78720bf8e749a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 260, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91efcc5af8dab70c42865e7b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 260, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c7c04444f50e40a1fe9bd0d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 260, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:551d132b7fe1957f1fafcd9c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 260, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f7e877788d4f0483fb831b4a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e1102b0e8c86b1205eb42ee1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30c32511b5b558094e40f4be:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ede71c7b809f014437af8af9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b241cca34679e508f80d520:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9bda0666e6bfda7d18f928e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9b387744271fb301791a5ca:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:88d44fd1fcb501ac14a4ca3a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b78e91b08f7c33555a2be8e6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:337befead30095bfd00497f6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:331a356bf87970533c9d0725:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:497348247c579141e0dbb7cd:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5f4b5af955c3668ce46b278:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e9fb0462fdde24b92c4ae16:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:397953eb4614419d1829c2be:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e1354f3cfd523ca8a9ad26e6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ae476711536eeb7a36e22892:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dfcfc043a7fa144ae814b82c:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:679f14fc3de034a6be2d0bd0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:087bc9b0d1dcdf437027e20f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1e23bc5db13a97e887d68e0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:73e95c4d2684383be003f37e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87d722451f1b7193edb0faad:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:721ebd866cffbc7ef5060c48:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd75add80aa1f487583f4872:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c32bba374bdb7e14985c345b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a021259694317bdba5827742:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df6a5dd9c406bb8ebe3d304d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:555a65463db746a00caa9303:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a13771e185ce550ecceb9495:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ee0e3af1db8eab565e6c4cb:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22fd55f789f1a68e66f445d9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:045fc472036e54cbb8252ae4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5ce09907c4bcffd15ffecda:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7f890a85ef1d61f9596fbf9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b44a56334a863cee1f199a4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:14974047485ab7c0046a98e3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a3b8ad471b72cb36bb601cc:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a89ad1dacc449a52d1ae6fba:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f89b9397e2f164a290b63b22:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c2006b576adbbdec1f51a99:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4265c05b34d7c35bd610aa2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c0690fd151f35e027f00ce7:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:43d345259f5826a942fdd763:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:305cbb6637f27234894cffad:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0aea7e42d966b3948d9e5bc7:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b67297d4c55629a3dfe7de3b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b055d9c44b56b5c3f4e70009:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29f944e58ecb95c0cab6e537:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35893d776cf8a9befbf40035:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a9c32c7625e434330e609edc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6128e861a918690c6e5de3b8:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69f0be365acf463400f838f5:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62d7d2ddaa30a852b22711d8:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e5252df8fd788ab6798734a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc1cc99f8445a13669f31a12:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a40e83beae3627e8be7e5e0:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:993de02d92df7259e27471dc:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11dd0449348645cb6381950f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2797e843c26bf42ebef56796:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b7c46a055482f10e1f7fc17:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a7264f7039f6beae0e9eb36b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:059b9e89e0f9a39e9ca07caa:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ccb8e1870088eeb3da0c8fde:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3cf32b9aab552f1682389315:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0324557b4afae51b6c3b6520:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:562d5fe09245bcb8558bc151:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d70aab02a83c9ca5067874b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e02e43a0c3c80a1601cfcac2:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e3f53db64493301d2eb03fc:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c68aab7c568da7cfe676b831:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0373c11329369971628bed04:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5931355aa7810ea88f612ff:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c8e6f7b20b98b0c3046f1e6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:44e50a716d69e952a59fe921:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54cbbc305b7176514941a66b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b89665d9169a60779097a93:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ed998e08fde27f01c80102b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:685e864104193bc3279ffcf4:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dee9853401a8cbdb728be055:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f08c439be56316b9080dd5ba:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51280ec5eb510c0a97afadb7:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab2ce9be62657279cb62c903:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:511e9000cd444d45e17ebce2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b912e73239338bc21b705ab9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb133102174bcb7528b070a9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d56d6b7e687b0ccb678de193:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ff07fed19d0c89fc5f1249d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b5bab804434183269879087:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:84e155c3fec8100bdca752c1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d54d43a30b0adf2a51d7aec:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a10059fe1b4256a9ebee490d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5d2835e70cdf0207eb971a5:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 293, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:86124d379be3e4413450521d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 293, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:efe142c7a8eaec77652d4306:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 293, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6298ca633d87a99776608f29:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 293, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3cd6558cf84d7a5994a6622a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 292, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d76549b65f6aafcc5670f615:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 292, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80febaabbac7480b645587ff:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 292, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd157bd523cb73feed988157:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 292, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c667fbd78f30442caeb4e97:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:08e62a44f1cc4060da22ea9f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa3baf8ac48d9ffb0b7ff915:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:929e09f46361fa24a970c282:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:64f46f6ac1a317cf8acc09e4:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:09756350b0ec681b3a14174b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:88ed4ec705e6a442cc3f587d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50bf33865a0deab0365f116d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:37af3284cc151a723eabe51f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0f112aa4f4a68cd7eba0c37:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:57a0ec0dad63bcf13206c0fc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65c9e08f92a9ce202c0bee44:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51ce8ed8a12ba504bfe2ab70:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c915e3987dc51091da144fc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:660306ea3c618a2c8d964ba6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51a70374061587647b530d28:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:323443dec2bfd8fc64c627c1:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea9b832e0022f5dfb2fb3c3e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bfe049b0a54b4bbe62048bf3:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07cc9ad2dbc6a87688551d5d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a47f17d3591dd007fe483f10:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23f5dfbe1d64b1cf8a522228:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:abe20fc8fce53fe9888b9c1d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5e60d2ddbb6468382c08734c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a80f96b87e8d723b92a77a10:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39919d49c9207c1e90f5a595:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:295a6fb411ad8c2a8ecf7d84:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23786cf4761e9b4b10e9cd4d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7daa4c4f395c335381c0d6d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2534b7357efb6be74fc48b7d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:798964b71191e9362595bc91:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d62d16366ccbd619783b8d59:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e727ac0c846d35db48a16604:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67cd960e6ca8f35adca973db:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a046ff2f9d09d7c2c33aaa20:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f8e25ac2ea9d85113b86073e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed206f27d5fd98e4879c9b15:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9aa1dbf07565bc86881ffc39:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d8c9dba76f0634f25015d92d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92d6713c5a1005b29650d9cc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:155f6fd14641a2072ebe821b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9cf6b23995764fd03ee7bea:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dfcf1c40c9a2804319bb0b0a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28654d3ec6fb4502546b014a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:03f2ae04e713ab4a0d75ceab:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fabfd8e25fd21bed19be5b3e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:63e1aae23481dce0a89b5b6c:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cad7109adbe1d0a97cb90b64:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:94ba63fc6f08206a978c264f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7299a5fb077300e87ab688d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d54ce421f99ef307f3f59ce:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61e2ab96ceb285917e2797d0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75aad09329a738f0363ccf49:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:de65158bf80d5a38663dcec9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f768955b02feec7cfa83e201:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2aa90e678f7c83ea667bfd4f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7835a32e046860a028c75089:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:579cf9fe496f94d8a2ee08c6:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed75ec25f215b211bfbed7f1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:536db834b5b818bb16daa79c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a36bf8099a156adbddd2ad7a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7f3d2e4a645d95404112b30c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c0d19e415b0bcb3aca49fb9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89779b3d6e87b2f15c2de4c9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d75e142ccb5692ea83e2aa4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8010f5ae742ae3fe2ea75565:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b05572e331ca11f13cb393f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bc5f6892031a4c76ee30c1b3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:53a373b84b65c14efd0a4836:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d789477e68b50e48e9a3f79:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:610f6229e6cda13d25a0b7ce:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e6a6bac193d16547aaf1849:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a41c002fcdd40680188ac946:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0241eb9082b557b013bab311:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bb89750127ac7be3b25d710a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a997f65f7e5cfe2a5a449f9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:84c07f9f155dcb320c251d6f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1691659edde0695028e9dedb:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39f0c8cce80e33915eba1af9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbc4c721c2d6cdffbd79c471:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5e5fabcf3cae0eeeaddc500a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18cc15219fd4a1b47cda8924:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa51313c6a41f885e17b07c9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6646e1db106cf350a236e6f0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a6b65f9da451095a51adaf1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:110e5571dfa4e32984d8bbed:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a13077063bb4b2e6453f258c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:de9461d82cb7e70d2efc9a9e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d6315c1652269686febb597b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65ef4deb55163817323437c5:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d07dcad3068d78d7e7caf1c4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1aad6d1c910112d41557409:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc8ca6089c697bc5a1492ca7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:545e4a8ad127d85f7be2b1a7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9324a60519705f2e6ec6f50e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0084135d34eab3a7ddb40dc:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c16e7b978ef83f776bc0da4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7384defad0b09055a7e5af93:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6e9827e301e4b6b4ff519bb:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ec7025b451dc2d96d979b36:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f57147da19294108965d0394:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38421e0749191c6415985d5a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a2d8204e9a88523541f8227:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:682351d3410d2bea2fa6f1b2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2870f64eeb1b539fb2f77edb:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a803e4987d80d76e1687195:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e56239e7fea917609bfd4e88:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f57263b7327c7058ad8556b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eaf990b375eb6070abebbc27:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:76f41f8162d567eb515bef20:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8961abd3696162e3ad8fa607:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67ace2dd0074d6e92a9d8365:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e243aa382a403df328cae226:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2530b95b77a385e6843b03d3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:084ed959138bf7612f20a416:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6387f001ebfa13c17ebd482a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ecb2114c65608c3b51e3a6d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:786534c9af01c99a755be110:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2284f5093519d1c18fc2035e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:882a6e804d22a30368d8c321:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c7750a82db748fcbbd9219f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:339a487d3651f5f4ea4d0b87:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47efb15edf8a829a6e32943a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:657b760eef76afe98564a8af:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ad7bfa9a4d549142e5e05617:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1bd284a9cdb1f1050c9a503:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c72896206e66d0f8765d8e78:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e86b247e390217db561a58a8:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8ea01478a316c176249ea340:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68b06f530c85fa183e07c6e2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4af77e879dbdf5782dcde370:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba8ce1ae3aeebe4f43d17812:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93ba3b56fec917437ceebcb8:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0dab6df61ca5ab3788ba126:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28c93c1f06a1d34b67e903a5:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81856ef4972db3d940df2c10:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5676985d18ae665f15ab884:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a36433584cc2c0e6be9185ca:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:709273537de2e62d3a52c427:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:654e7772810f847d66532c54:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f24a3617fe02c91abbe5041f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2350d22c125a175fb10ca667:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc69b1c49aad3a88ccdcb851:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e233250fb2d6462001431874:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ba78f9f3c619329d795af5b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f145d6756d87026f8b6195c6:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58792b538006a2dac978a42d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cec5bed52d30e8e43a4548b4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b74b7c08ce380e74701f600d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:510ab59890d54ef432234f49:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b67ea8d1b41d8f2d5e8d54b4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8451388bc93fda2c517d8782:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d19a66aface410f373b064c0:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c3c4963d436bb3972057ac4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f3c306cf3ae7ebe371984e1b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2529c991ceddf93e1774ca1d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddabb0077ef28acca27ca222:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a2d479f2af661bb1445796b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8cda16d57cdc7e27c9476e6b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf5b4d14e4dcf0727374be6a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b93fb286759b8e62292784a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:13353909aaacf548282639b0:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50d51c7ab2491e5f0bb3c92d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0366d43cd88b910555d9988d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee12a2ee10b48580d32943e3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7af336f36b376a6c77ca6842:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f05e1a2045f80e1d836c1581:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e71edd437a6a8a07d177b13:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f62ca6a2ed225e51db14225:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a5914416468517a0b0ff145:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d676b530ca5ed1ced22a0387:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf91d881fa5b9e0aaae9ea52:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62d6e02278abc36dc07ddab2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a017efb6234aa7b31af7d4b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:76ee37061a5405d319505fee:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:78fe39463e18193f6f722c20:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a5b7d707e00b38aea06af53f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed14cbcf4b9acf1423e88f23:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b8b1c5bef3c12a8de7c85e3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1aabc92cad14a9713c13a5eb:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf77734c437918e8f7e7731a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:74ae31ed6c8172c695484630:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bd9842ee9d27d9fa47f8266:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e3b45aa18ab1b92efe48ee5:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fab8bb97fe11f1b62e8b9b27:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a098a7001506748903e34244:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3bd74fcf796dacaca3c1424b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b9afd90d8b5580a4cf108b7e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdd5f0a1e2ca707225df9d2c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:05235228d48d2c0f6e0a02f1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3827ae0d3110802458c785c2:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c1cc7ddb7b3ec2c242a9684:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e203b12dc1b7929794a55556:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:908f2ea15cb1a58118ee639e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e3f7b7be4d54ef85f042af0:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ce606021e7d267fb211dc32:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2fdeb459ae6aa66bcb33f4fd:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:71ecc43061baea0bb8d6f17b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:605a84bc61b79647bb6befd3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a3e0719675d7b131349bda7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06dfab5f3586102c06b11bd8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e155e533cd3b40514de4cd2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:914f5793661bb16fba025e11:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c0614da34fe6bb649a6bccc:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1301ac3ce27d89bdfa895e5e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdacdc72817d55ab8b915247:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89cc9329d953fff42dc5511a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40c004c56e45ec837a9dde20:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8deb92b9ca849dc07976c082:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3cc0967c1105c10f9282b781:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1043ad4e1629659b9f22e73c:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:193473b8f7115a21adf5d14a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:86a1d90c6d74157ebf00f674:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ca5f56a805d7310b948bf64:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2fe55000dba0e9569483234:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06ffb824478d7940d6e87cee:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c74f3747bf72aa58fcc6362:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01278efe0ba68e26071eb7e3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b9c5c5aca5f7d162c33ba59:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e93cd98533b0ad7dc32081f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e986e02f7b5ac320d20e4db:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5bcabb6356565e568c2385f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e8fc97779ff6fc27ff2d45bc:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9e35df41f0ce4b91d15516e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a96974765448c82f9782697c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8766b052f48a9b2141f028ed:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df72184ed05171cb9855dbbf:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e110b6cd6d99c2b269d499b1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5129ad65df8c55a84df23027:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92b3976d280a542f9c9aef34:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ab53eb1f38044b52bae38c3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:422f9eaa4fe7c1f8a50ea2dd:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b7d11b16b688d51b539206f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef3060728bca1d4cf0f019a9:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8859b39384bb3674cf6be6bb:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c0555aaa0232ddd41d8e2cc:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3bf1133b5e36177f94ca1a0e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:707c76f4fc5aae7650e4e4fc:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:449fd5d4a4819694c23f312c:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fe60882766a2777ab2626e30:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb046d267656bd328719ec33:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:991ae9c245f3648bc6bb555e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fab016da4b8b42a7c1518243:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b065ecb382c44810956dd656:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e140c312293fc636872e53b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9157ea8f673273b0215af17:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:570df5aaa332d4a68c18bdf1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:488dcd8e14dbe73940196dc6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9f9f6be71a35fb5486446d3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:acdf1c88b4e2bef4ab7c7849:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18d213a8bae7b3bccdf613e8:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51c409d6916c6d030ad1637e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5e4f8f1c65e2a9b22d376cb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:812d78c2eb35aa32c37a597e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d0b562dbcf35f1ea7a09744:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:27afb6de88ee9ec114ec0486:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb078bdfdc6ae4277678c522:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcbb32bdadd494ebfadcfba8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d8952313ac103008ae74ad1:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e98e88936d9b8ac9043b78f8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25415ab0085df742e5084d07:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87c192cfcdb1856fad3478c3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11c13a2daff536edc2fbbe36:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f10410c1e300c2ea71a05cd0:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d2803789969cf8d09304ba37:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f5547636f20990e8dca02ee:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d95b5182df9b6f9aa3613891:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68cdf01d844d93287c153277:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:60814b3b9638b09ed683e08e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e8e8d0c6bc5136bc4ce5ede:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f2b89d0dbdda52ab6d6e50ff:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4473de6bfe4b7235222e5fb2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d56af286301a9a66175f01a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f413a7105a191f14f3fff3c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9edb6cfbdaa5a2c6de9aefd7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e59f601468f29e99099e93e8:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4f926b73ab2335f7487d643:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50af27c07d00484fb5fdc2a4:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bbb7d0605872b41c51f51a7f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b93a7bf8525b63edf9ff956f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa9ab6774d73d5d7f080f96c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c8678b7d2c720fb65344a6e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d905e9d4b72f076d33c13d7e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab6e78fe497743f0c665271e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e299c8c41ebb74c6de1b84e7:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33cc9caa2c248519b00b234d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f1fd043e4bbabd2b700e5c8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3717b8258bc68faa8c07dde4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e10ac8eb3e3daed71a532fe:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a77ad1f41898319f0632a08d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e54fcfb4598d53919e450d2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0405e987ba41594ccabf7c73:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:775e102d75596b0afdbd90db:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fe6a66db0a05664fecbf76b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9358d6026a16d685e434f6c7:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2eb3e35689abf90e8bcbf390:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:64996692610d2a5b58ae7d13:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa647d7e2e1d8ea274a812cd:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ddd764f66f874821732ae62:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc4d6270a2dbe0133403c5b6:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:628db00b8e82c227f96717e2:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb14ae09fc1f5737d251e397:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd5bca27b03de8f6c3efb042:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:512a26d67ae11e8d2c860f82:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81fd2f3cced5a25b0c96e652:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58534746f5ab71a986a774dc:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d64aac749500f9e584897d5e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e2863d514f130ed51f903779:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b72853b473ea0de9384cea26:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e0dbf6be9b6a683d91e6cf6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0cee46a52059780eb66ab99e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06790038d3806fcc21e42a9f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ff08fbabe8d5f4a1c81ec57:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:531c9d02df53393befafef5d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:169b51ac2e4533c2440c21e5:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc31237f03923622616a839b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7283b6ec02cc568c4f5eb14:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fed65e8f556072b24338fd36:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66a6d74d15970f5e754937fb:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c83eb465fb0b7fd14c9277d5:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e70d82a90ad462e10db6ba3f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d897d668231ae80605f98de:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf23e82bf53e897e161d8837:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea4c7ab25c8f096b2dae54c0:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b278426dec26bc256dda4e75:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36eb18bf9f8cf102395d747e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c0a4cd9942e23c9e23c0a05:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8653fc3c4700ab9b3b7649ed:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c8d0b8e1bc9b24f82d07acf:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1b26ee01196e1ea7f07ea3c:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:04c467da98d7f56ad3c018dc:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a6a32cc5d567d9009b1f49bf:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58eb17212bae51b923a77c58:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b8594591afe004fef731dc8:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6c5b9ef3773c7e70f81d46e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eec5eb8d0f252bca0b15e523:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83e304d67f5a1171ffc5a562:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2707bb2a79bc6e792efa79fa:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:522c14a39f5da560faf1a97b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34241971057b390340995342:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a41798edf526abeeb95b5363:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25f1f7e791ea0799b23ae83d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8420346b0e5a12ead19bca99:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:083954c9631613f6a35c706b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6af055c4464294cc4f119f3c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb7fba40afa7b36a7cad11d3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5680dcedb9099c4244e49510:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2838d4c9fc818e92adacaaca:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3803ff8f185cb6c0b864b625:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e2263572d15ab9c42991d2e4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6964a194108140010a727229:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdd5dee44a8368a4fc90409d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5732eb3e29a444c6ee65874:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7faa9327efffa3dc14583067:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:443b0c18c2875f8a8e37a1f8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:77ede63aebd73d60183893a5:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df8448608a68717291c0eb18:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5b49b6d4d2aa5d67b4cf944:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6634979ba0ea6ec8bbf26fba:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e564ae2439e35d264ee76006:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f4e7ed27454a23ea024d707:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:264cc44adff3e9f1fdf86480:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6bc0ef43c0ace87980e8503f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2307e0ee9f6e7a69fbb647ab:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fe71e38c197f8e48fef8e48:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4cd42555038fba0422650601:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98c99c306daeb5bf3af9094c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee2489584e1bfc6af99707a7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac323cf2eb1ee685dfa522c7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c33dbf8ee24c8a70b5d50045:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd558108f5d2fbcab5bfa70f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7006d38bf86be5d8a59c9742:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:523595510e7296ebb20acf7b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf39e2bf39481de8f10256c6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7f413f388cfc14001fa5a824:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1d6b769c01f031cf30a8b17:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:728be71ed26f414e2a0ca7a5:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:45ae094c037d5fe0704671e9:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:891e1b52cd7ef73071aa5dcb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd3ed5fb89e2c15c468a3fdb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1299983abee789571c9ac385:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f72357ec9461fa7521d7c9b4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b86a8e0188534972a6465778:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f685189d7a3f0217b9ed8fc3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40e67a322acc3fbbefcdafac:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8355d71eb417a7a94fd9f0c6:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25620227f87474990b837326:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba24955414be2a069685756c:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1a8ac0e48da71859fc97ff0:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9806f7229f492f708ead3da9:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c3bc051e04ce799f18b5654:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eedda693c65651a283918434:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ae764a9ca349c92d64a86586:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93f1ce93859b7cc9c5d8a81f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b975b51c04f7f4696fd7920:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c89642c2a50b9cf7f18e6002:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4148e4dc8b3a5e56083fbdc4:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbdaa8face17ddfe1a98b26f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2eb0e0b393805ef04c67907c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4fa8143ffc721f6280353a0e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa26d6ae17994438f6c4ed9c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:72a98a6bbb53a9f2fda94f60:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b67e800ba2a3e0a6eb459272:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35cfa778c07a952c2859155e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a43da79e448fa36e76ee3d1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:654ce403dd7e0f274a036ae1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a0020baa4fb513918b26339:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6254bf5d53b4be7132d2d910:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e39ea51203e2b96e7dd12341:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a55f8d812b545a871dac869a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d43103db8078f4d067ad3a1b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ce3cd2bb718f5276f6680a3a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4f342a36cd26224a43d1bb5:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:099e0c7417dad965c8f917f2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2582f7bd8825d9c4af4801d2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:03afd18817e9b25bfaa5b698:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9aa0c14bf7183cf4d1872346:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:317c4491ffba7f46d902c29d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d261c4848f9b77d95286b9a4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:edcac26c71b61ec69705eed3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ec66f54076ac77de1d51dd52:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c634002c241dc6fc71bde51a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b11664acf4121adcdc67111:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a4f737bf5b5e0901b4538c6b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:82c075aad51ae5a7cb6af5a3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35c16e6d9badab247444d9a4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:407a51e349e7ec8b16b3f1be:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f845d257f2f3270388cbde8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d48825df0e1c4ac8bef6ec00:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:912ac89969d3782530f92967:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1047e0607d8e4830663d0505:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:14a876cbff1015cf972308d6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:64018ec56babe0609839e5a9:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7253fcf0764a46c7ff28248f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0103dd7c32ef82a1beee95e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7a3acac21a9c4228a9ab0a1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:836f0a0387bf6119f675f5a4:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:902cad43390a52291cde0b94:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6a1e6a0e0a2182d6f2c21c5:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d8b2d3b72fd44649717b06ad:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e049a91a38149f16d87f919:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab090279f03bfa45f1375e79:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ece7ae3f8666ed6a9184b183:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cbf0340aecd26990e0b8412:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aadab2875c22b9b2c62403e8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 554, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a0aa98887d61798dd7a853fe:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 554, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6b30f34bddfe17b78849f021:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 554, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ecc2b752bb2631f83cb9f6cd:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 554, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12401d2736944514687c8ae7:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:48eaa4c95dd48f7beebf6fc3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cb7b1d761b288a522673f3e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4899b65e15f570ab23fac6a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f846081c095ca7beba167a7:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7bd9a399059a2078b58da8c4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:afafaa1fd9f895ced2c4c1ab:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa3b3cb3af6810aa3a31d11e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be63bb38f41a4941d553d5a3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 555, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:717d8b1c969ea380f52afe36:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 555, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f73f895a308386d08adab8ca:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 555, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4768dad76749b886aaaa609:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 555, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fbcf752aab364fbf97acf0b7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e6ea4f9d07c0de6c3691949f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e51fdeb1a3c3c840b7dac1d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8bd8787839eff57d85091339:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d918756e233ac53f641ac325:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f19b01bc274a489cc430614:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f459fd3966f63f9bd9b08692:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6cbcd7834565e3ac20042685:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f178357f12791b2cdfdba3bb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c52b7fb5554667a02023fce:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:63be0db1c2badf142dc7235d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b4651e4262601b752ea2da4b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:00e0fa31fd378d209ff756bb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab1ce7b7f1b4874030601e80:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f1398f7918a12f8c420e511:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdbe56852c5cddde7ede66b2:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6c933cb93ef12640810048f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cc5bf01719660c9e3942d1d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c66bdbc23fbc7bac9b650e74:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85dda89cf9227245bfe7946d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ab0b66376a9bc06310642de:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1329356437645d7c33293744:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5bb04b7b365ede2a6b923669:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81200c3e10f262a29d7f8abf:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38b12c8ad0af9d6083610899:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a9fa002546c887a1e6052fde:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:838844fb9c237c11b0d67901:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f74612b21e17f245b6789018:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28a4b7f5c29ba1e8701d7a6b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8bee79293ea193f85bd8f408:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c864c098b1f1dd7239a1bc3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ca1aefea5abe484166e8d84:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c79e74d67cd41d9608cd6b9:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a428124d2572b4ddc5f78b69:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6d82e1c9f5b568f5f96ba9d4:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f37d48c368380b33e429322:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9f2acfb91ca30ae535d98d8:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3db0e5eb4d0e8fad5ac9bd76:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5c5c9d6fbe0078b0b3c45b3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d62dc2b322644af0c4b1433:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6bf98150b40fc673ae6f2e06:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:050eb5903da9dc589640c7ec:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41a75f7a84558f7ef2a91791:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:786721715802ac6831afefbf:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c588bf292216b33b8b19ec91:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:573b8a5483cc63b744071e42:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:77cd96f9eb526d87db0426bf:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e60a3473dbafa2db4894f1b6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15a0a79ad1804a523d4fdcc3:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3aab322e68bc40267fc81e00:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0930f3bb683da1de730e12c6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9dec1b50aa908b6697c50c8b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:982a1f24d0672f692f3ab260:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9f263e16582be7736c59e60:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4eac891687aedc6e20a2fe4:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f5967e9bdb058ef58331e86:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9aa936effa81886e4f8d2aa:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e53af6ec6f48e3b6ef308c30:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf9e10a49a97e9e66cbedf88:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6b1658a665077870054689a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6415756613b00d63f603dc5b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a1d2ee67666e9e76a433d98:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65d2a2a53f5eec3803ba833a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1aaccc663a04aa8228bacbc2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:355688f1d93ce79dbd3551e6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40e09f3769968df2baa702fd:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b8c53ddb9c12a4fe9446e92:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c87101a4925e298aa6ebbdd6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15690dcfcbeb905525378eea:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3269f1c6715be32f5918b43d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81289dcf882179350140a4f4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd769d1328dfa811f037b9e0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5eee8c15777f36aa41d2432a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5ad4f998350d94b43e5cb66:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e6ad48460feae23637d2cda:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34c9876ee6537c7f90292664:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16347b895187d24626199382:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3f4f12649bfe560792742ba7:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a546f5f90b7b7ac01e1a4595:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab7f7d5f4f3c302115c82baa:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8d8c7d1aab200648dee1ea3d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:310c7ba7ac4719062817d591:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:afd236258599e4a58de17166:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6d80797c7424475cbb46519c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d524de9e4d11c6f16c1a4037:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc3844f44d7e3bcc429c3bc6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e99aea1e810e1e7b4c305923:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be6c3ff25d7cf8382b90f9dd:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:acd2a3c99c6af78c6e465be6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d7b6bb08c570fe4af3024659:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e11d7dcd3f35cc8663b8b81:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0bdadaea6ebd39a130e72e47:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:26a1681e7f60bc341323040a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:82203c0c43bf32f0d93c943f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c2e44d61fea9c77aa16d3be:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a9cee5dbf3a0ccad88aadfb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5d095b12fd7472d3bc8bb69:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31b9f493b908961e6811ab9d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:edd0e2ee17dd2da478eeecb4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b455ae9104c57566d6564252:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:640b059af09b84e9b7978903:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3247f454153089665ec9098b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e95e6a94af634c7109c3f246:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62b69142175d94f3fd4e2d72:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f04a0e1c43053b82d55ecd0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb9458a3614f91257f274be5:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93fd401e3dc8e507c8a1464c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a4247841838fbbc31bfba19:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a98804b3a9a9bb3d16bbe974:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:279040d0da84d719e1008a7e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5013d3f84a766ff90e1d66d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:95a8f41f1a8abd2d3a9eb1a5:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28f66441157133b175210bdc:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2480546c3f0750ab8889063e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a6bdab83a4b7ecef3ef647c6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c72598e032df002a371c721:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed030c57b48d07a26fe3fa94:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f49d1a1b8732118af3e1be90:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41337ab27205460f2133b691:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:569e1b7cb9d952a5879b5568:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e40525b20d565be581c4ba0c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5359bdc648fa50d4cac1891:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c29a5edeaf7b44106b0b134:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c367aeeba3ed47df08cd733:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:866a5c381de4ac62118dff85:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23e4fdffb6c15673bcdfa30c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf34ab76605a322f3022c5b2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9dd8042ecdbb399e6c0049f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:465eb1946dce6dec3a2a5ac0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01d94f0b21a5dd2c818b3cd0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38f2113fc226de905ea3e002:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f67e5938b2605fee54199d36:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f9fac3a5c1238a6bab5e8372:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6978b8de7ca7784877ca9c1:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b4ee082c0adedfe3e83ab0a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aaec28883f5106f66e678b38:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4b666673a2b69725f3dfce6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1dea05de416e4b30a3d9b3bd:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e3b4aef7ed11d9fe77e2363:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:539335d4d742aa862a840934:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbb73448f12660f057c255be:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69f213a32868f74881039073:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f61de138da767b3a06938b2:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e34402918bf0bcc3917c54ba:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8fc95d8e56831fb80726a1d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0809c5a5bf2a4a227982b51d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ff22df9dcf793090dd745cc6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28b2c8e61c631a7343e23fe0:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a000537d614295fdf04d0998:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:131b6f3346e6a783bda21564:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdf794b73f61d7e3a877b6da:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0a5494be9df8886d8899ab5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:255ccea5a91c2b6559fce709:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8078bf5f14220044ed9caf41:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5798b04ac0704290fc794ed:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:49dd3e24e9d0ecd0bd6ca487:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:db25e3fbe8afafe5ede4a0f6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c28d78407447277ce418d6e2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b52a2679fd113557ab42323:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb116adfca97f9e016d66bec:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7047e796c12bd31f40ca0d40:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f03be949c020e94742f9e77:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d25222ed9c8a3bd46b12cc20:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:463bd53f0721e822217ba46f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0ae468377d98732c59ce2f3:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:46479f8f0607160ac6dea63e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb2996746d4a3a8542f6ff1d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:077f815d1122414cb8415a85:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d573dd064a37b7a0c385b8d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3412ff2ac67063d787d1c17:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:462618b61651570b2846c938:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f1183ebd02f47a05a9bc908b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e73459046bdc8819f7836a29:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ae6647d9b35a7839e4ef83e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67f1d172f24a2b42d4d9abb4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fff63af9fd2bd2bca03d9d72:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:637de0b4ce60d431b3fbe6b9:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4748d544ff79abfe5fab9847:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5a5948642f75f93c2f3cb77:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac5c668d5d11be78678d4a84:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:126562d789deb80356468d41:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:57a2008078f742ab6c8626e5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a4553b77d5b53b5790e3d20:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11a4cd6425ef50288dd89dbb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c32ed72850524b6115259475:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:43b0d5e950a0b165bb220323:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f8fe2a27635c879fb8628e97:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ec1668afb0674d887d30b5cc:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbd521c43201e26268e3aa7a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93ce0e674e3a155e92ce6ff0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea556a3766a137afff9d5143:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1c350beb0696f45d0b636d0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a561199f0e30a86517d4e77:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7c26cd5a928391332a449a6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e112c4da03e89806f6b51f2:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d21883d41d3b3b5c5968af26:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b9715bdd36b25a38ad6486a8:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:820626b763f8825dd3068332:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93d08bb7e0b93cb303105554:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4a68ccc2d5e311494876e75:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d24a7c504ed2622641cdbead:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 283, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7cfdccb7443a93a634c433a1:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 283, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:817b57139b63e108f8790fdb:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 283, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d941b724bf9df118dd6a0aae:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 283, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ee3c0f4347f3f58e953611e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3ffa2df372e8f6dfb0d8c38f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:471380d453ba14d8f9c03077:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5387a48bf735045df5884247:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2082df371acd1246e6974878:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e51957d703aa7a21a9126f7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0b27fc187d5b92f2f0baf02:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:388029ea4c6b9db0b05498d7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:596dc4bbff62dd637bf1abe5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f95e6852ac5ed8170400816:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d53ad7b791e6754ee684dd46:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6c5abb646277c869e8f594f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a0e820d5880fd053a23b8742:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e24bd548aa26f5edb96ec44:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4020df98668c9b44c45f1f60:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a8f4fcc01b49eaa3bb6b30d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c8cdf2cb4c1de722f45f3936:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea686bc3ab05831b29e7008c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c938b97813786d8c82237b3:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d649dddc9b41e2ca653328b1:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07b6de7eb9bd7361c486b345:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eef957bbb536e28fe4381bdb:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:19a7d6b2323b098d9e22032b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f2aafe130fc41373b6640f18:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7611787ac567633d94de55ed:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:648bbf706e496cd9f6908242:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0cd71dfc18adc305ffa5780d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:262929de5f678d99fd81523d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1cd8bbf14e4a59fe63113214:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4402a299cdebbaf033c9050f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f882ed980b69673ee6a7b58:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5d8322afa755ad70d59d60f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b959e9bf77a5adc5a30d2ce5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d189f2df0711b7fc2c771265:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c16014d87823dd357d1456d8:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:128f395ac37d7006e4e4c367:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c29119aae7c02211285f7082:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:697e27e971a1806c494ac95c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8047d9de8466175cbba13672:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5025ba4651167be2693e5f6a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f63abc827e81e246ac37e965:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a80ed0d47a0d3026ed1dfbc:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e7fe1b673f6177b0fa1dc76:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b1a6bb2f1ecd4b535f679a07:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a50961b7fd5d31f3923d1b7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:441340c226bb1935b121e41c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67f031bfe4e9e92d197132e1:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e0babcd3043ac68136a197e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5e495780c777233199b2eee:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8bd553e22df1683db52d5e78:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4892ab35e607d1b27a490c56:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a5f71abe56130d9b6d00a9e8:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67be1599e86d96b0a79d4b3d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c8656d105aa907f177cd2e4a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:501c446c78491b8f0a6b7edc:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:db5889f39922205969c462aa:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a67f9b5b6935e1e6b2573e9c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5b299de7d39edda771cb7973:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:342af4ee63b0cb125b271a4c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:625175dbc051d9cac87f6110:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ac053036472f06ebe164a62:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51f4549cfb5680161d097026:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6231a699832f887534551fae:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ec98f845e66e9a7221f2312:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f53aa74c3869112597b5fd30:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b13b8e3553d6936297b72580:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92cd14a08b1204b3c59807b6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d250f37f1770efbf1655af6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7bb180e591734d11d40b8605:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e7adbe36391fe00ae9a4164:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a73aabb2b12d1a995f9bcc66:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5978de20565039401c0f9ac:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:814557a8578a6d65feb6e938:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35e4e09a850e616d48db8bfd:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33b52c598887d9d5905859f1:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:488ceed82540659725d8c36e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23c3f398425fa7ae16f59f98:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b7eb7993227f08865ceba7a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:851cc9d5fbda6c0c718c4fcb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:111c2d37ffdf960b57346949:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c79b07c72f1575107a85cd9:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d41feeb008c6f9c5b2bacc21:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:896b4776fa05a557779d30c4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6ac74dc2de34b129cb9155a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:86965a2f8f90ff782961f060:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cef8be7cfcedc847fc2cfbf5:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5469358631eb0c0fba97e23:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31afd4f6e7c73256fd123f75:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50e9b66e55411fec32bf17d7:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e8b47dd39a6ad545fd84c1a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e915c93c80ef2f02d4d2c4b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8ff178390db9980396b19358:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b6008f12b6d461b6f5ad121:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:55057bfba949c6ff0663fd0a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fcf6ad681a5145718215a021:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:428aee950ad2683c813213a0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b91f11fcd38154517858dc2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8259f446982e956a0ca11695:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d36a15fe27e140b6370c6f7d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:637cc81e55d6b29b3859c430:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8baf32af0f2efafb0884ef1:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be4b7e8f2f86f141c2d8d3d4:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51275459581e4b1ef3b00b7e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa2a4a1442ae92254da80c38:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbc1dd0f7c711abf7adab9f5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f8bb96e5dbcca559f3dec2b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4abdb09fd01605d136e6fb41:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:806d0448919413ca550c11d0:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8be2af3bc78bacaa2c393ca1:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ed67e4159be3dfc4fbe7482:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35fd0ba5d43edd45fd72ccd9:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9bd3410efc8104bc569c0c73:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f7fc6e2d78bdcaa8ebb70d1f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 588, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3ae94e3a3615b606c6b0faca:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 588, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:253b1f5b5850bacf600b161c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 588, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5ce3b5ce3aa46d05142aff02:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 588, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ce51cdfd50612eeba18aa9b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 587, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0138be8aae7f889727aa1d1e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 587, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b7e759881886577dad273fd:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 587, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef6acc71d3fbb16066661281:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 587, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:27d8757c9da7f3430e5e0ced:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:194742a1a19aa36e85a58c7b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31f3f29c21bbee6875a62524:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d7621d7b8893eedbe393db0e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4a3eb5c21fc375b3e82b339c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d10cb67b66d9c0fd81ac7e58:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1848c23cd4d35898d022bdc2:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a30ca432414e10d4fa53896f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6c30c88fd3483ab14dbaa1e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16b40e61859c5318b4586afa:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a0c1d6a2009f8c9b6913994:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a0bab24a3ae166837ddf068:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40ab5b0e677bf9770e81090a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:78f00f7c7878440a2410f788:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f15638d409cc44fb52c8b7e4:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1131e5a097ac7e10af06232:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c2b532fd0a1c316f6a7827b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:def4366ded6a6df896671e94:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c25d1c92537c287c8d1f9ae:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5bb8c0d305fcc9190b60bd0d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3529da347672ddc6efcd7cbf:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3eeb11ec8c5b578dd385f823:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:55f8f4450093cca5395337b4:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4845c32b3bc94002c70ccfb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c29d1885fb3e711832008e6c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:501e2691a84e6ddeeae2f9eb:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f293099110937a1af731b7c4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92488c2f6185ebce0ad9f995:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a07bbdfe3071266e2b3b2ca:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6595a8344e600006d2123f9b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c4a4e147f7b1588ae0bad92:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4979e947e876782bb7b71363:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd3f7c9444d9e11b215839ce:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:866d5c488eed87d63b94c01b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b60f0028ea5cc952f66ae471:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7279558be989c68b7c5da9e7:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ec625bc4ecb9c56c9fa0cc5:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e4a7f385594c24fb238ae4b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6f2292977793d8d3b0bc2a0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e331f261aeb0da55b58b86a9:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e4af581b0e4f9fad15de3b2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:102959b674a96c013936409a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01f5cf08902258a28cdb1212:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6cf9cc6e93265277a8e3ed0d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f221f191c4d26720283f4d1f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e11a42fce820771511feaac:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15450d6b94feffbbd82fe3f3:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9e0925630cc8108b1dbf766:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6b98f2b0a9e5e50d70c9e675:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:241aa15d6210643b37819f49:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c8b4b84f80a8b426b676149:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dfec15d15538d2133f64505b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:332c305d9f9064cca7c91b1f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c9d3d9f73d49180fd4e70c5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8353849f306c99b801e76cbb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e5c16ccc37f603aaee2e0e8:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5af1d3651ed77b0e4f433bb:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7b1cfff32e388417d93f789f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f3fc69214f96664fd12c2681:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f10ee0a7b7901b7b0cffb768:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b1460dbef236498d111bc421:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1776662fd2b3d4233b7b1f84:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bbd1fa1192f750484e786f94:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8ecd72c6f944eace26d5a99e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a5dbb5ddf4c974b338731c6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2afcaed2770a429bdfae7681:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f0fd919b00922f0fb8004de:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91493ce1dfd1c9efd5084510:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c214eb63a496e85f4e8f0331:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e9860c60f475cb0c1bbc1ca:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b31bc0c29ed3b7f06b10dbda:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c73bfb4d052e52de134e6f20:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c31a0ea252ae23f905fd135b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:305a069109d09eb613b4548f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e45ce175a6959ad79175b27e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:656489c3fe1025652a510de3:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91f5ffb49f77fc333df264ed:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66e0e51ffb6f3943c3015dd7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c12252b469789af7f5190a79:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:563ecaa55e9433ce2d40fb0d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6ef55093396c3904f96ed65:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:924e4b96ca5999023e35d70a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b8e9ced855b770c0618a521:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bfd687a18b43c9f071ab97c9:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0902d059933889080b0053c7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee9468b38861c78cbeae36de:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c77559f45064cd0b43746d8:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85244a117edda6d6143e169f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65530be55a82ba6c9e4e0535:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e11bf957181a9d4a1bdca5c:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6bb216eb099bbe2cb415eb1b:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c2a1c77b0784c41b332a688:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa2b06bff614037acff1d512:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e9a16be72ced2fe22e4fcd4:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7df43449888199a23a8d880e:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d7067b9c047e1fc87bb947d:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8f52a5c304a45097f71bf97c:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:63d57a7aeb929060cf51188e:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0ed1675348397a71ec61d8b:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e4f1cfcee5f63659228e6f6:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b53b47523823292cca2a20f0:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 801, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3bf115a8cf096d0947013ea4:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 801, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b0199d15491378b758402f5b:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 801, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:572f0ba1d3ddaa3fa4ac6f91:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 801, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1e0301f1ecf0f1847123e18:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:040fc74e25d32f5a1706d9a4:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12fef32144a2172110026e00:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:249c8645ae2858840c0d1c78:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39b9d4042a63f6ee6f7c2466:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac51c42652a70e790e7111e4:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23dd18b5e052cc51e1ba4052:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47f13308a31f26e9f85206a8:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39cfd2cf1e0faef8f4c0c05f:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:10589703550d91e77b38af7c:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5ff0bfd990d8f473981610b5:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e3d2edc1d02debf8b325bca7:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:90ead2aa54f0e923806f9b7a:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6266d716e135c92d2241af78:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:04100e1e91887aecab329285:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:640310b1795eba44df0405f7:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:823ed2dfef942950a0fe052d:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89e8e93ebc09ab77a2d12ee1:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2f03a7e02d4c7a0b37a5c12:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8df3da6709753e818e475063:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ec0ca15717c6beba10df0969:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34e19699b6fbeb555cc4c3ec:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2498f981bea7f1ea2f94f562:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1181ce3fb218aa0ae26cd2ba:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80ada521c29562b2d79f27b5:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4d3051156c025b3e9c13127:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b17e8036bb65af3443e673ff:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:858226223a62dcae09bbb797:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8be9a77dcb5da3744baf8878:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75d8ccffbb0e46bdf88c8778:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5eee2c4c2557caa27de5ade:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25092695911e018291bee049:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:249580a78256b5fd65542bbb:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:297c6e026ab9cf238b3c1d10:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b91dd50087e2e74714d9cf1a:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e56f2ee1a2eadd1013f0e3e8:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:942cd8d852efe24d6c5a4dd5:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:334d03453308ec3c45e12e23:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a267b801453dda4069a69b25:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98025ad456cd47b9148284a1:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:90f596b7d7aa2159085c2b63:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:244f73488ae8b18c97afc783:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15eb4c3530366489aea10f56:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f0ea261f90dcddea760524a:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:db92f0e465c6399c5c4abec6:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:af3cc83d4f3621aacab12778:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e2293e4fd434a600da4d0bc2:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd59a2e1e889f95b39067f66:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4997c1961582643cdb3747f:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:197e69acca81ee3796ea18df:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b58c019eddd3d2c33f22959:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ef468fb72ac4cb64a81acb9:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:78eabe08f32e049a8cb4f8b4:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:48823570051dca86c519667a:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:27243d43f64937e1eb9b625a:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a9c17041a8da3f8c51d28132:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f47dcdcc829ef63b4c6c5d0:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f4459f112d4feb1760b4176:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:de191bbf708498615876e957:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d76351a595e7c5c928de2ae:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06c1c1ac6665eca225c2382b:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fe46a01198229bfa376dd10e:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e064899b758fedf86fa2005:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4318d2e14a9b1ce3c16c2894:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f066a2f4cbc8579afd9f1982:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c33d8efefc0fc8118834420:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3bb9967a166f17cbcad10f3d:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9627aaf17a3806f0aedc0837:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a252aaae386b04fc387afc9:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38edf84bee3e6ad9c57f203e:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c962093142dea21d0ec3dfd:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ca03f0754d5ce00a62bf0ca:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:196f3a6bd50e835d24bda265:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2eef5b738aa0b07684bc1dd0:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06f6dcb92ca8f002d5047baf:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a477e556b4abb5cd224ffc4a:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b4683cbe2ce2d042e8341c67:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:05da9714c27d666caaa2a782:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58ef310a92cb4b8b9b6a34ac:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:539c270926beb6b3742b164d:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fab0fb117c6ad4e0061ee617:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6df3c133fec6415c08fe51dc:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:753c524e8accedbf5235e6aa:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a255272ba51fa1edc1aa73e:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ddd679c6df6a0939e230453:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b23c7e2e9a40cbe74d0d0aa:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca673d0199005992cdffce6c:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:031d1db0d38c30f30506a4bd:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:45b3799fabd153c39f7b6531:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8f743a364012858f43563b93:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9229baa933c2fe26956e93c7:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be37856b3136494b863c28fc:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5b1cfec010f249f02c487e0f:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02a0a62903ef89c49f64ee06:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7dd94f0d108bbb8627c3b1f:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8467efa8c94001e51b138cba:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:045010ec88340767d1b89b2b:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:13c6045871929ef1628066ba:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36d14866217fafab8eae7780:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e7ed7df6d8739bb28b8e150:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b49b03914fdf279abf49926f:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:141a5bed53a942c0406932ec:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea6882511a6c6fdde770fc68:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c50529b6eec075502f964315:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:44bc47aee0dd3387580fbbe6:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ebe5cf8672d9330665e9e010:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:538db869e56e44d7178b3be8:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c841d71a8f7ea07ed3b92208:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5db15071608cdf33b0e85961:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d69cd166660da942fcd5aed6:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:97127e2f57c841d95a72d242:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33f2a2f998988627ddfa3f98:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + } +] diff --git a/variants/events_brier_seed17/tokenizer/chat_template.jinja b/variants/events_brier_seed17/tokenizer/chat_template.jinja new file mode 100644 index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4 --- /dev/null +++ b/variants/events_brier_seed17/tokenizer/chat_template.jinja @@ -0,0 +1,89 @@ +{%- if tools %} + {{- '<|im_start|>system\n' }} + {%- if messages[0].role == 'system' %} + {{- messages[0].content + '\n\n' }} + {%- endif %} + {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within XML tags:\n" }} + {%- for tool in tools %} + {{- "\n" }} + {{- tool | tojson }} + {%- endfor %} + {{- "\n\n\nFor each function call, return a json object with function name and arguments within XML tags:\n\n{\"name\": , \"arguments\": }\n<|im_end|>\n" }} +{%- else %} + {%- if messages[0].role == 'system' %} + {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }} + {%- endif %} +{%- endif %} +{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %} +{%- for message in messages[::-1] %} + {%- set index = (messages|length - 1) - loop.index0 %} + {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('') and message.content.endswith('')) %} + {%- set ns.multi_step_tool = false %} + {%- set ns.last_query_index = index %} + {%- endif %} +{%- endfor %} +{%- for message in messages %} + {%- if message.content is string %} + {%- set content = message.content %} + {%- else %} + {%- set content = '' %} + {%- endif %} + {%- if (message.role == "user") or (message.role == "system" and not loop.first) %} + {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }} + {%- elif message.role == "assistant" %} + {%- set reasoning_content = '' %} + {%- if message.reasoning_content is string %} + {%- set reasoning_content = message.reasoning_content %} + {%- else %} + {%- if '' in content %} + {%- set reasoning_content = content.split('')[0].rstrip('\n').split('')[-1].lstrip('\n') %} + {%- set content = content.split('')[-1].lstrip('\n') %} + {%- endif %} + {%- endif %} + {%- if loop.index0 > ns.last_query_index %} + {%- if loop.last or (not loop.last and reasoning_content) %} + {{- '<|im_start|>' + message.role + '\n\n' + reasoning_content.strip('\n') + '\n\n\n' + content.lstrip('\n') }} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- if message.tool_calls %} + {%- for tool_call in message.tool_calls %} + {%- if (loop.first and content) or (not loop.first) %} + {{- '\n' }} + {%- endif %} + {%- if tool_call.function %} + {%- set tool_call = tool_call.function %} + {%- endif %} + {{- '\n{"name": "' }} + {{- tool_call.name }} + {{- '", "arguments": ' }} + {%- if tool_call.arguments is string %} + {{- tool_call.arguments }} + {%- else %} + {{- tool_call.arguments | tojson }} + {%- endif %} + {{- '}\n' }} + {%- endfor %} + {%- endif %} + {{- '<|im_end|>\n' }} + {%- elif message.role == "tool" %} + {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %} + {{- '<|im_start|>user' }} + {%- endif %} + {{- '\n\n' }} + {{- content }} + {{- '\n' }} + {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %} + {{- '<|im_end|>\n' }} + {%- endif %} + {%- endif %} +{%- endfor %} +{%- if add_generation_prompt %} + {{- '<|im_start|>assistant\n' }} + {%- if enable_thinking is defined and enable_thinking is false %} + {{- '\n\n\n\n' }} + {%- endif %} +{%- endif %} \ No newline at end of file diff --git a/variants/events_brier_seed17/tokenizer/tokenizer.json b/variants/events_brier_seed17/tokenizer/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2 --- /dev/null +++ b/variants/events_brier_seed17/tokenizer/tokenizer.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506 +size 11422650 diff --git a/variants/events_brier_seed17/tokenizer/tokenizer_config.json b/variants/events_brier_seed17/tokenizer/tokenizer_config.json new file mode 100644 index 0000000000000000000000000000000000000000..9a87e3822403fb40a9524bb75edb66c6f0585750 --- /dev/null +++ b/variants/events_brier_seed17/tokenizer/tokenizer_config.json @@ -0,0 +1,30 @@ +{ + "add_prefix_space": false, + "backend": "tokenizers", + "bos_token": null, + "clean_up_tokenization_spaces": false, + "eos_token": "<|im_end|>", + "errors": "replace", + "extra_special_tokens": [ + "<|im_start|>", + "<|im_end|>", + "<|object_ref_start|>", + "<|object_ref_end|>", + "<|box_start|>", + "<|box_end|>", + "<|quad_start|>", + "<|quad_end|>", + "<|vision_start|>", + "<|vision_end|>", + "<|vision_pad|>", + "<|image_pad|>", + "<|video_pad|>" + ], + "is_local": true, + "local_files_only": true, + "model_max_length": 131072, + "pad_token": "<|endoftext|>", + "split_special_tokens": false, + "tokenizer_class": "Qwen2Tokenizer", + "unk_token": null +} diff --git a/variants/events_brier_seed17/train_log.json b/variants/events_brier_seed17/train_log.json new file mode 100644 index 0000000000000000000000000000000000000000..231ff75499630f5a5e909e23644788c4579fa428 --- /dev/null +++ b/variants/events_brier_seed17/train_log.json @@ -0,0 +1,964 @@ +[ + { + "step": 1, + "phase": "full", + "loss": 0.47767234593629837, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 1.299889337271452, + "batch_question_ids_sha256": "9815e3084b0060ca1bf3f53e09caa2e8ba2ecfbfa18cd5a90c1b6ca1dc29fbcb" + }, + { + "step": 2, + "phase": "full", + "loss": 0.4745664047077298, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 2.020658330991864, + "batch_question_ids_sha256": "0766e00bccd0b9d63535226a2a7454d88d591d6a1adb32befcd64d40174bd87c" + }, + { + "step": 3, + "phase": "full", + "loss": 0.3339335899800062, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 2.7444814248010516, + "batch_question_ids_sha256": "3eb8fb253ed8d59b98fd13fb12ae3aa08473ff98bf07f26c7537d8d4c6999318" + }, + { + "step": 4, + "phase": "full", + "loss": 0.30728491954505444, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 3.4489749232307076, + "batch_question_ids_sha256": "b72b9555459b9660b8cfea9c061b0d69e21b213181cf19973fa9312c1d740e4c" + }, + { + "step": 5, + "phase": "full", + "loss": 0.3180481344461441, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 4.1599197732284665, + "batch_question_ids_sha256": "851370879eecc2bb2b872bf9f26a90b0c422dbb910faa7ab99625462dd70126c" + }, + { + "step": 6, + "phase": "full", + "loss": 0.31872232258319855, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 4.87426084280014, + "batch_question_ids_sha256": "a44ac9fef40754316a48915525e55ee3c8c5f613d2a3a206a70c53f6a360d6a1" + }, + { + "step": 7, + "phase": "full", + "loss": 0.2807605154812336, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 5.582096985541284, + "batch_question_ids_sha256": "dd15d5df027d50d8bc83cc139ce58f36adf2613954b88024320afffaf4f275ce" + }, + { + "step": 8, + "phase": "full", + "loss": 0.3625841060420498, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 6.32407046854496, + "batch_question_ids_sha256": "b16e21ce00030d172555d2ed47f3935798c5293130655f2d4bb114fab8d4f82e" + }, + { + "step": 9, + "phase": "full", + "loss": 0.1536292051896453, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 7.142426532693207, + "batch_question_ids_sha256": "db13a8f636762d4e5c56f60dcecdc4a04f7acccee36ed4846fe242c110813e33" + }, + { + "step": 10, + "phase": "full", + "loss": 0.4065534844994545, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 7.96108222194016, + "batch_question_ids_sha256": "d29a2ab7bae9372a6460e6d41d2320b0a13b55633fcc21e2a06f06a64a2b8f22" + }, + { + "step": 11, + "phase": "full", + "loss": 0.2941920473240316, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 8.778865645639598, + "batch_question_ids_sha256": "e435c49e7cce06b5c387742c1cf2eae1fb3f57349a03ff3229713bf8e9179b7f" + }, + { + "step": 12, + "phase": "full", + "loss": 0.5255483770743012, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 9.569397198967636, + "batch_question_ids_sha256": "52067a44724ec88a3f85cace2ddba01f511bb396cbf27877697e85efd4f0c68e" + }, + { + "step": 13, + "phase": "full", + "loss": 0.18976907199248672, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 10.306229672394693, + "batch_question_ids_sha256": "d01e765c3bf9b2b1edf10d82d3b784f51514a21682d040badd6d9c1c29c3069f" + }, + { + "step": 14, + "phase": "full", + "loss": 0.30221015959978104, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 11.033171433024108, + "batch_question_ids_sha256": "f7d76860a7e7abbd079278179b2a6054aa8b1bf90c6e38db91f414cf906eff6a" + }, + { + "step": 15, + "phase": "full", + "loss": 0.3888988718390465, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 11.780395239591599, + "batch_question_ids_sha256": "1d089d97e3e6836ec1771dba1776d38f6a8ca414f96cc3c25fd0053ed3e27a44" + }, + { + "step": 16, + "phase": "full", + "loss": 0.2125754952430725, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 12.491722567006946, + "batch_question_ids_sha256": "abd204307499e62673daa334cd3a4533448a0ef971b517e7acbf23783905ae91" + }, + { + "step": 17, + "phase": "full", + "loss": 0.2280114609748125, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 13.19165222439915, + "batch_question_ids_sha256": "6d4029fa8044653ff0ccbe80e4371cccd55b6fc7c6462fb81d77f565bde1e2f5" + }, + { + "step": 18, + "phase": "full", + "loss": 0.37986432015895844, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 13.910702037625015, + "batch_question_ids_sha256": "88ddecb0e43b996acf8e321896454ceb6e79095126d68380cfdc7c74ba66d04a" + }, + { + "step": 19, + "phase": "full", + "loss": 0.3772949390113354, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 14.622391032986343, + "batch_question_ids_sha256": "3ba3b72125f7617c5cb4dfc031843bc10773d8aa7747ef6eef76ec813abc7a11" + }, + { + "step": 20, + "phase": "full", + "loss": 0.47917328029870987, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 15.332007034681737, + "batch_question_ids_sha256": "a8fbfc400a3dd61e2468ddde908d8e6914e8742e66df75fa209575bbd49f674f" + }, + { + "step": 21, + "phase": "full", + "loss": 0.2207504389807582, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 16.107440183870494, + "batch_question_ids_sha256": "49a7b1f6bb6891cd06f7c495942bbdbd98cfbd7435f49d77922ab364b9c1c5f5" + }, + { + "step": 22, + "phase": "full", + "loss": 0.2577735111117363, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 16.850639042444527, + "batch_question_ids_sha256": "d93e27a20a5666b8fe2ea5489dac0a58f99a770bfb0d4ee6870d36621ed78dee" + }, + { + "step": 23, + "phase": "full", + "loss": 0.36209800839424133, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 17.691069215536118, + "batch_question_ids_sha256": "88845418c486e4c7d856761b8aca680a7ca52ec59c84219e708745af76a445c2" + }, + { + "step": 24, + "phase": "full", + "loss": 0.2882579481229186, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 18.40779590792954, + "batch_question_ids_sha256": "b7fcb8d686d0a87ea4fd6ded60d595c7ccc4b132fc739e047564d72f5944577f" + }, + { + "step": 25, + "phase": "full", + "loss": 0.269174687564373, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 19.135179361328483, + "batch_question_ids_sha256": "b2e6af990ea722e6cb276f57a736c23766306ae29b211bf00447d868b434c962" + }, + { + "step": 26, + "phase": "full", + "loss": 0.34036025404930115, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 19.86950681824237, + "batch_question_ids_sha256": "67253956a0835f33db1fa53443fed54476a027338e0ffdd7726dcb48fbef1c9a" + }, + { + "step": 27, + "phase": "full", + "loss": 0.33926529437303543, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 20.584741784259677, + "batch_question_ids_sha256": "b37cd333ba07e45e13875704fb0cc42b2f44fc5317737590bb27d88a7ce61452" + }, + { + "step": 28, + "phase": "full", + "loss": 0.3321617506444454, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 21.315836602821946, + "batch_question_ids_sha256": "d2a5a95e218ea89d7469dfdd8ff320de68f88ff52171675c0ccf7c9c3d95a891" + }, + { + "step": 29, + "phase": "full", + "loss": 0.2353851981461048, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 22.03636402450502, + "batch_question_ids_sha256": "1c6aa40b463139c53b7a93df96c3a1db9c53fac3123be7e8dda4fc8f53bc7996" + }, + { + "step": 30, + "phase": "full", + "loss": 0.11482883733697236, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 22.730047752149403, + "batch_question_ids_sha256": "be09ca5b88c61e0617249157b2d75545b12e403abaf2ff570be1e6d30856d742" + }, + { + "step": 31, + "phase": "full", + "loss": 0.41507381945848465, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 23.457241214811802, + "batch_question_ids_sha256": "08f4544d1dfa624e78df3609c0059204a16146f8421bf85c14d0b10b5926b448" + }, + { + "step": 32, + "phase": "full", + "loss": 0.330725422129035, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 24.159707821905613, + "batch_question_ids_sha256": "f8a0c95d4ce5a21d3b7bdd381eb7c9d15a251da10b33f0a8521a3e81eeb1450e" + }, + { + "step": 33, + "phase": "full", + "loss": 0.2639482840895653, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 24.896084915846586, + "batch_question_ids_sha256": "4c396fcb15e0a31d52d00473dc2f3bee3767da23146fb5038b6e63d9ad5016c8" + }, + { + "step": 34, + "phase": "full", + "loss": 0.26776977452391293, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 25.612479377537966, + "batch_question_ids_sha256": "c59ffed718f7ded62b1028779c65e8759de67070b26019e8642f75dd2e43a512" + }, + { + "step": 35, + "phase": "full", + "loss": 0.43445201084250584, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 26.31534642726183, + "batch_question_ids_sha256": "0e54ce4db6221fac98c15ce77d50701153708881dda86adbcb2776637215c9bd" + }, + { + "step": 36, + "phase": "full", + "loss": 0.5263467519544065, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 27.034690125845373, + "batch_question_ids_sha256": "24c257b893015e139aaec823c9fa8ed630a7d2df239552f229ebb832a7746723" + }, + { + "step": 37, + "phase": "full", + "loss": 0.32995557998947334, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 27.750695827417076, + "batch_question_ids_sha256": "bc22d73d0ac85948fd9bbd0813992b093f9beb230652c6234447dbbbfb4add35" + }, + { + "step": 38, + "phase": "full", + "loss": 0.28109178878366947, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 28.491634809412062, + "batch_question_ids_sha256": "6e47f7951fb42ce413ed2f4ce54d8832b534a3e9d0f2e35829a1608286227ef4" + }, + { + "step": 39, + "phase": "full", + "loss": 0.256468391045928, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 29.180634811520576, + "batch_question_ids_sha256": "b5bfa6d3b4f6544cb42bc8b82b7c88ad3430367ce2f47dd0c3182d7f55e26796" + }, + { + "step": 40, + "phase": "full", + "loss": 0.2841529995203018, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 29.919747012667358, + "batch_question_ids_sha256": "e356d8692763396b1f305f967bc42bbea222b3adcfd7355c8ffd489d7c4a8702" + }, + { + "step": 41, + "phase": "full", + "loss": 0.3729940578341484, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 30.638639828190207, + "batch_question_ids_sha256": "eddf9bafc965ad35c8a9189e6a7b19eb336e5f2736ab490e5d0f7547f885b883" + }, + { + "step": 42, + "phase": "full", + "loss": 0.20839229971170425, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 31.362661720253527, + "batch_question_ids_sha256": "93f507828c7fa8715902d87edef912b05fe2920c25f400745b5d4b1835ff1fcd" + }, + { + "step": 43, + "phase": "full", + "loss": 0.27495690272189677, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 32.06642537377775, + "batch_question_ids_sha256": "50f72e83de9b1b9daf154f013aa536b1454d8d8c13b503ee6efa9007c721faa6" + }, + { + "step": 44, + "phase": "full", + "loss": 0.48349061608314514, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 32.812288109213114, + "batch_question_ids_sha256": "71067cb1ca0d666b5a16e53227ae1840e468c0c39e879ef8004ac73520b396f9" + }, + { + "step": 45, + "phase": "full", + "loss": 0.19684232398867607, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 33.537176745943725, + "batch_question_ids_sha256": "b9270e713e4fca02ecc7205c7ac1e320d3c87286eb7237ed0b02a4686b626bc4" + }, + { + "step": 46, + "phase": "full", + "loss": 0.2759946044534445, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 34.26172788254917, + "batch_question_ids_sha256": "fbecc4d2e806b5b42f4f82a7e99465e1fceb5b7b61dfc3cd167654deeef897d0" + }, + { + "step": 47, + "phase": "full", + "loss": 0.22207287698984146, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 34.97563951276243, + "batch_question_ids_sha256": "9fb1e3f8e3eb222e0e377eeedad5b0df90d11389465803b67a9b5090d2e79ca0" + }, + { + "step": 48, + "phase": "full", + "loss": 0.2445692487526685, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 35.7044816268608, + "batch_question_ids_sha256": "8ec776449fb8174bd55b829feff0711ca000a9e09d903e741da6f91e2d21a136" + }, + { + "step": 49, + "phase": "full", + "loss": 0.09901037975214422, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 36.49502827972174, + "batch_question_ids_sha256": "a20669c6cee5e157b8b54729bb88921f4ff6df89b4dda467ec25abb10f0e78da" + }, + { + "step": 50, + "phase": "full", + "loss": 0.14326033811084926, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 37.1994747761637, + "batch_question_ids_sha256": "cab2a2f34731e8a959300614046a8b8fc4c926971dd5909a1a6e3de0ad3acb26", + "dev": { + "questions": 372, + "objective": "observed_outcome", + "eligible_questions": 372, + "target_ce": 0.5348790703109285, + "target_kl": 0.5348790703109285, + "target_tv": 0.30709812289697785, + "by_target_kind": { + "observed_outcome": { + "n": 372, + "ce": 0.5348790703109285, + "kl": 0.5348790703109285, + "tv": 0.30709812289697785 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 372, + "ce": 0.5338139083905662, + "kl": 0.26381131578303585, + "tv": 0.20872586212331487 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + { + "step": 51, + "phase": "full", + "loss": 0.3137478260323405, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 43.16678562480956, + "batch_question_ids_sha256": "46758e1c6996a3f1a932a11e8372539949f8ac8f4c0415b36f65c713e195096f" + }, + { + "step": 52, + "phase": "full", + "loss": 0.2717863286379725, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 43.874052685685456, + "batch_question_ids_sha256": "d87fcd81d136c657c13e49aa033db193bfbb121e4a721fa86c16aad4c0a7c39d" + }, + { + "step": 53, + "phase": "full", + "loss": 0.28308067098259926, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 44.6021969569847, + "batch_question_ids_sha256": "4fae217aef8823502570d85ef34334cdfe2994ba567eef62990a60ac5e3ede4a" + }, + { + "step": 54, + "phase": "full", + "loss": 0.6188238859176636, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 45.31346751283854, + "batch_question_ids_sha256": "edbcfec53ad977392c69172be6901bdb4eee1725bf8fd83582f64ac6320e77cd" + }, + { + "step": 55, + "phase": "full", + "loss": 0.42757737822830677, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 46.016700938344, + "batch_question_ids_sha256": "7af8e2cece43ae333d4687839255ee144e5968fc9aa063efa2ff6dd96c42d1b2" + }, + { + "step": 56, + "phase": "full", + "loss": 0.30507429130375385, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 46.7371188858524, + "batch_question_ids_sha256": "6d579036d49f401569b6e29499a444725fe7a64de0907bc07bd389fd352e615b" + }, + { + "step": 57, + "phase": "full", + "loss": 0.23919871263206005, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 47.45829302165657, + "batch_question_ids_sha256": "848946e51c61b43a984fc49047cb8c9f0fae9f28c6447d8b4ee2c74dafe198bc" + }, + { + "step": 58, + "phase": "full", + "loss": 0.3340693097561598, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 48.18450126796961, + "batch_question_ids_sha256": "7a207cb4e3ff6157802e3b560ec0b63a488a8ab0eabaa5bcdea45eeeb107d3da" + }, + { + "step": 59, + "phase": "full", + "loss": 0.3097129948437214, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 48.893560361117125, + "batch_question_ids_sha256": "6ed68210f517203266a035e4dc739a5e91ad1773fb37b2d954aabbc97ce44cc5" + }, + { + "step": 60, + "phase": "full", + "loss": 0.2643849514424801, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 49.75948201492429, + "batch_question_ids_sha256": "cbce464330f46f2f323539fb9d7962cede14e7074be588012e1dfa684881e6d4" + }, + { + "step": 61, + "phase": "full", + "loss": 0.22575335949659348, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 50.51938905380666, + "batch_question_ids_sha256": "7212e07d655e85ec9b791cd6cde65ec3e9ebfa5224d1854c76d9a8e98cc07be3" + }, + { + "step": 62, + "phase": "full", + "loss": 0.3371951300650835, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 51.23941683862358, + "batch_question_ids_sha256": "768e8960dc65d1f0cc4b13cc2c90dbfe8db2c685cceb3a8ed4f834e59a13a2f5" + }, + { + "step": 63, + "phase": "full", + "loss": 0.4498938750475645, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 51.95728688593954, + "batch_question_ids_sha256": "ce0e1bf7240822fc1f373178c20604e9dba0409aa63f2b51415a5f4db012ea78" + }, + { + "step": 64, + "phase": "full", + "loss": 0.3990790694952011, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 52.69237087201327, + "batch_question_ids_sha256": "6b8c351645894c06829c8dffb3045f3a3081c0fcb0b24c924b9576410c78d47c" + }, + { + "step": 65, + "phase": "full", + "loss": 0.15212762681767344, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 53.410304352641106, + "batch_question_ids_sha256": "06de39fb8b5de5ce866344dcb10447ed159378857789cc67daf23061f43e3e7f" + }, + { + "step": 66, + "phase": "full", + "loss": 0.25516666239127517, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 54.1299799894914, + "batch_question_ids_sha256": "1728e00f1c3d6bfc533f2891740148ba51b40dcb766709fed2a21437282f6ee1" + }, + { + "step": 67, + "phase": "full", + "loss": 0.15473376447334886, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 54.84718861337751, + "batch_question_ids_sha256": "26b40bee228ef411ff536ea5c58c9a2fe4790fcc250d1ec97719f5c2b6051983" + }, + { + "step": 68, + "phase": "full", + "loss": 0.5024608373641968, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 55.56789685878903, + "batch_question_ids_sha256": "ebcd0c069c1aad6db15af178bfb86b95232c842c724cae991202c127dddf4df4" + }, + { + "step": 69, + "phase": "full", + "loss": 0.40726556070148945, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 56.28566949442029, + "batch_question_ids_sha256": "021f795640c8ebe5b45f63378c3fb38ba2c849f5d70b747f653482f046de3fba" + }, + { + "step": 70, + "phase": "full", + "loss": 0.3981996960937977, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 57.01161473803222, + "batch_question_ids_sha256": "968367da65a754ce0d7eb72abb8cbb64b558c86aa9d8adda10e119723caa2bb9" + }, + { + "step": 71, + "phase": "full", + "loss": 0.2851921394467354, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 57.74261846765876, + "batch_question_ids_sha256": "5f2d81f647af85f4a48f30464c10a61c6a85b7d5d35f687d527ade870b0c1e7e" + }, + { + "step": 72, + "phase": "full", + "loss": 0.517420694231987, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 58.47589023038745, + "batch_question_ids_sha256": "6780ef3f330f6cd45a377b3f3c501ecc3ef2ae65d08769f207f4a68c152e3dbe" + }, + { + "step": 73, + "phase": "full", + "loss": 0.22744558565318584, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 59.20974231977016, + "batch_question_ids_sha256": "d521c674f97a615228053a262b84518189f9580e2c73b32eee51bcef5ed76d29" + }, + { + "step": 74, + "phase": "full", + "loss": 0.31863199174404144, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 60.064713726751506, + "batch_question_ids_sha256": "6c26d6bf906bf556525d65a2709f2f59a7cee21e2f62c8bad8c7015fcde12f89" + }, + { + "step": 75, + "phase": "full", + "loss": 0.30666476860642433, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 60.784851206466556, + "batch_question_ids_sha256": "afabf643cddd0692428615c4aa81345acbbcdab1c52a5c9a48c815f9e2700755" + }, + { + "step": 76, + "phase": "full", + "loss": 0.3913169279694557, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 61.525260652415454, + "batch_question_ids_sha256": "ba8ce0da2f777532f12642355dcce1c4d29f8e0865010f26b4bf03c56d7723d1" + }, + { + "step": 77, + "phase": "full", + "loss": 0.2860779892653227, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 62.26424302812666, + "batch_question_ids_sha256": "9e0a4687654abac2570a41f8674ab95fcade2ac8e5e238d4661499dff7d58834" + }, + { + "step": 78, + "phase": "full", + "loss": 0.34327320381999016, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 62.992976791225374, + "batch_question_ids_sha256": "e36ae8ba0c5eaffb1a534d69421fb9118e4f45453e90b83cbbf5f29ef284cafc" + }, + { + "step": 79, + "phase": "full", + "loss": 0.39875966031104326, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 63.6979681737721, + "batch_question_ids_sha256": "171d14c1b2acfe9895a9353fe3dc68dcde6a998182a601a2746bf8fb10de20e3" + }, + { + "step": 80, + "phase": "full", + "loss": 0.24940739758312702, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 64.45239959936589, + "batch_question_ids_sha256": "5091eab21841dd9384edd2c719867375591caf9d084ba03c065878eb451a5674" + }, + { + "step": 81, + "phase": "full", + "loss": 0.34674376994371414, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 65.18292573653162, + "batch_question_ids_sha256": "f94cae69521603c0b60eb1cc6bb5370d2ea3d089f5b6bfd9a260b41a766ff974" + }, + { + "step": 82, + "phase": "full", + "loss": 0.24664756562560797, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 65.89143331162632, + "batch_question_ids_sha256": "f804d3278fd4930c24d09af7cc2e1fc70170c35cac45cbed9779d59f2d765e08" + }, + { + "step": 83, + "phase": "full", + "loss": 0.242635240778327, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 66.62519173696637, + "batch_question_ids_sha256": "19b60c4f93e1ba7368b1f066c22c4ba32b8b66981562b2cbe2485da7433c07aa" + }, + { + "step": 84, + "phase": "full", + "loss": 0.24826265033334494, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 67.33410876244307, + "batch_question_ids_sha256": "ebdf0b28c90afc00f6691d1a09af18f9f66ebaf71377d0ba210fc4892869014c" + }, + { + "step": 85, + "phase": "full", + "loss": 0.2863457463681698, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 68.0715844547376, + "batch_question_ids_sha256": "4f04bf635d321e106f9856ba99a47fd6c2c033ae0450aa32c6dfe595a74ae4ed" + }, + { + "step": 86, + "phase": "full", + "loss": 0.18521394208073616, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 68.79707364458591, + "batch_question_ids_sha256": "ba9c22de6939f9c021221fe5a43ff5f50b15a70b21ed4effe8d97caf79b049ca" + }, + { + "step": 87, + "phase": "full", + "loss": 0.18474309891462326, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 69.51761453226209, + "batch_question_ids_sha256": "c9f35e67c5374cdfd4ecee262a8723f0f688d0387f073072324c9d259d9f4323" + }, + { + "step": 88, + "phase": "full", + "loss": 0.3266881722956896, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 70.239103092812, + "batch_question_ids_sha256": "64ab5437c436a5c934bb05696c9b73f14c9838d3b210eeeb1114438303808400" + }, + { + "step": 89, + "phase": "full", + "loss": 0.16573298256844282, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 70.9650420229882, + "batch_question_ids_sha256": "6ff06537c7c8217ebac7a64762ff5ee340f367a0f7652d25d5cd799c0d15cbac" + }, + { + "step": 90, + "phase": "full", + "loss": 0.26366132125258446, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 71.69922364503145, + "batch_question_ids_sha256": "8fd09d36031bf8ee878da3b38a0694c5e416636fcd74e603f03dd205a9a3ab33" + }, + { + "step": 91, + "phase": "full", + "loss": 0.2215642500668764, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 72.42362531926483, + "batch_question_ids_sha256": "322fb22dffd39fb986e5bfd20ac78a9963d2d5b01978598e3ea9ff285c9b09d3" + }, + { + "step": 92, + "phase": "full", + "loss": 0.2945997603237629, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 73.12151438836008, + "batch_question_ids_sha256": "d0417dd44fb2608169bd7f2ab17d57acd203be7f8242f3c2e89ad317009a2e56" + }, + { + "step": 93, + "phase": "full", + "loss": 0.1663272229488939, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 73.81322857644409, + "batch_question_ids_sha256": "57be67994833e0a92fac3ba6c7b6b1fe3b8383d53284136aad0b9a926f82aa1f" + }, + { + "step": 94, + "phase": "full", + "loss": 0.17362246103584766, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 74.53701240383089, + "batch_question_ids_sha256": "9c3cbfa9806ac2c618d0a1e3ca3ef41c91b3a86c8aa45a36071383124ee71262" + }, + { + "step": 95, + "phase": "full", + "loss": 0.23721570521593094, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 75.25338559411466, + "batch_question_ids_sha256": "100fb8785fb8c78e98988f555871d3d9a4ed2716c9bff93597e8a64de5e2db46" + }, + { + "step": 96, + "phase": "full", + "loss": 0.19234935194253922, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 75.98261929582804, + "batch_question_ids_sha256": "c781c81a212d6bc7153d4ecdfb49ff5f6570c3bde61d9f7babaa983e433d7348" + }, + { + "step": 97, + "phase": "full", + "loss": 0.271314961835742, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 76.7082808027044, + "batch_question_ids_sha256": "8b19c5c7751e07c22469266bfa40bb1eaac4670fe203a8c3e8a289350956893c" + }, + { + "step": 98, + "phase": "full", + "loss": 0.346242094412446, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 77.43025597464293, + "batch_question_ids_sha256": "043573f3a457c1f8cfbcefabfef11e430a1dd573e779dd7e19f7dc26a4a1ebad" + }, + { + "step": 99, + "phase": "full", + "loss": 0.2261855776887387, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 78.14046130888164, + "batch_question_ids_sha256": "933b002848b57b1ae58768d4d38da9823ca19226389dff7c41f6ba3e58ff14b6" + }, + { + "step": 100, + "phase": "full", + "loss": 0.15166849168599583, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 78.94848389271647, + "batch_question_ids_sha256": "2a2b48544fe4380e81979729c21f00410faff377b407c4afdee1bef7530a71af", + "dev": { + "questions": 372, + "objective": "observed_outcome", + "eligible_questions": 372, + "target_ce": 0.4421807345122086, + "target_kl": 0.4421807345122086, + "target_tv": 0.2684592128219396, + "by_target_kind": { + "observed_outcome": { + "n": 372, + "ce": 0.4421807345122086, + "kl": 0.4421807345122086, + "tv": 0.2684592128219396 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 372, + "ce": 0.46751975308439736, + "kl": 0.19751716047686577, + "tv": 0.17476332997830224 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + } +] diff --git a/variants/events_ce_seed17/backbone_config/config.json b/variants/events_ce_seed17/backbone_config/config.json new file mode 100644 index 0000000000000000000000000000000000000000..b815be77563f3eef88baf4b4c8ae59706e618ad2 --- /dev/null +++ b/variants/events_ce_seed17/backbone_config/config.json @@ -0,0 +1,63 @@ +{ + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "bos_token_id": 151643, + "dtype": "float32", + "eos_token_id": 151645, + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 40960, + "max_window_layers": 28, + "model_type": "qwen3", + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "pad_token_id": null, + "rms_norm_eps": 1e-06, + "rope_parameters": { + "rope_theta": 1000000, + "rope_type": "default" + }, + "sliding_window": null, + "tie_word_embeddings": true, + "transformers_version": "5.17.0", + "use_cache": false, + "use_sliding_window": false, + "vocab_size": 151936 +} diff --git a/variants/events_ce_seed17/best.safetensors b/variants/events_ce_seed17/best.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e15630e96caf7b07c526825c2098e7cc962606ee --- /dev/null +++ b/variants/events_ce_seed17/best.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:582c01139e29807b5f2690335271cf5647583cbfc3dd963d5967792bb3963992 +size 2385039280 diff --git a/variants/events_ce_seed17/config.json b/variants/events_ce_seed17/config.json new file mode 100644 index 0000000000000000000000000000000000000000..ca1ac62410fb271904dd5f06eaf598058e67ff70 --- /dev/null +++ b/variants/events_ce_seed17/config.json @@ -0,0 +1,69 @@ +{ + "input": "data/scaled_games_v4/events", + "output_dir": "runs/events_ce_seed17", + "model": "Qwen/Qwen3-0.6B", + "revision": "main", + "init_checkpoint": "/home/rwang/openjev_codex_20260917/runs/v3_teacher_coords_multi_seed17", + "objective": "observed_outcome", + "loss": "ce", + "reward_samples": 32, + "gradient_checkpointing": true, + "set_head": "attention", + "steps": 100, + "head_steps": 0, + "batch_questions": 16, + "microbatch_questions": 4, + "max_microbatch_tokens": 16384, + "eval_every": 50, + "max_length": 8192, + "seed": 17, + "backbone_lr": 2e-05, + "head_lr": 0.0002, + "head_warmup_lr": 0.001, + "precision": "bf16", + "disable_native_triton": true, + "validate_only": false, + "self_check": false, + "schema_version": "openjev-decision-pipeline-v1", + "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "initialization": "local DecisionModel warm start, fresh optimizer", + "data_sha256": { + "data/scaled_games_v4/events/train.jsonl": "4cebf698e4b71dc37d4d0e985b6577dd565ac370ac82b6f47f375aa6af69f8c9", + "data/scaled_games_v4/events/dev.jsonl": "0c352ee1064a5fb5d2aacad59cbea57afb897886d3d09e5ca108c866a5f893db", + "data/scaled_games_v4/events/calibration.jsonl": "4a8a8a2b81c3c46df6bb4c8690e97374dd0165cbef40566bb7f87c1f425d5562", + "data/scaled_games_v4/events/test.jsonl": "884bec71891668199925809d8708347f56f7bf06e8363e4a436347967b1423e3", + "data/scaled_games_v4/events/ood.jsonl": "3c00d845524634ca3b34559761cdaf0c52b7b7633d5517d1e7ee7f31697cea48" + }, + "deps": { + "torch": "2.14.0", + "transformers": "5.17.0", + "safetensors": "0.8.0" + }, + "gpu": "NVIDIA A100-SXM4-80GB", + "parameter_storage": "float32", + "forward_autocast": "bfloat16", + "parameter_count": 596250498, + "train_questions": 1124, + "all_train_questions": 1124, + "schema_counts": { + "records": 2356, + "questions_by_split": { + "train": 1124, + "dev": 372, + "calibration": 368, + "test": 364, + "ood": 128 + }, + "eligible_by_split_objective": { + "calibration/gold_distribution": 368, + "dev/gold_distribution": 372, + "ood/gold_distribution": 128, + "test/gold_distribution": 364, + "train/gold_distribution": 1124 + } + }, + "loss_description": "per-question proper loss; complete candidate normalization", + "rlcd_candidate": null, + "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection", + "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing" +} diff --git a/variants/events_ce_seed17/summary.json b/variants/events_ce_seed17/summary.json new file mode 100644 index 0000000000000000000000000000000000000000..4871142b47cc15f2dcdb1af8da65a10b58d563c2 --- /dev/null +++ b/variants/events_ce_seed17/summary.json @@ -0,0 +1,136 @@ +{ + "best_step": 100, + "best_dev_target_ce": 0.47277030621569244, + "selected_on": "dev target CE", + "objective": "observed_outcome", + "metrics_by_split": { + "dev": { + "questions": 372, + "objective": "observed_outcome", + "eligible_questions": 372, + "target_ce": 0.47277030621569244, + "target_kl": 0.47277030621569244, + "target_tv": 0.26819442928540077, + "by_target_kind": { + "observed_outcome": { + "n": 372, + "ce": 0.47277030621569244, + "kl": 0.47277030621569244, + "tv": 0.26819442928540077 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 372, + "ce": 0.4926936688108985, + "kl": 0.22269107620336712, + "tv": 0.1714566637835109 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "calibration": { + "questions": 368, + "objective": "observed_outcome", + "eligible_questions": 368, + "target_ce": 0.5122280796241342, + "target_kl": 0.5122280796241342, + "target_tv": 0.2856829761147523, + "by_target_kind": { + "observed_outcome": { + "n": 368, + "ce": 0.5122280796241342, + "kl": 0.5122280796241342, + "tv": 0.2856829761147523 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 368, + "ce": 0.4798722818150309, + "kl": 0.1935759627343198, + "tv": 0.16869317512676552 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "test": { + "questions": 364, + "objective": "observed_outcome", + "eligible_questions": 364, + "target_ce": 0.46045028819154266, + "target_kl": 0.46045028819154266, + "target_tv": 0.2658522378700697, + "by_target_kind": { + "observed_outcome": { + "n": 364, + "ce": 0.46045028819154266, + "kl": 0.46045028819154266, + "tv": 0.2658522378700697 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 364, + "ce": 0.46643101807363885, + "kl": 0.20566021774752138, + "tv": 0.16303083589195544 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "ood": { + "questions": 128, + "objective": "observed_outcome", + "eligible_questions": 128, + "target_ce": 0.33238140555610585, + "target_kl": 0.33238140555610585, + "target_tv": 0.23786875248045813, + "by_target_kind": { + "observed_outcome": { + "n": 128, + "ce": 0.33238140555610585, + "kl": 0.33238140555610585, + "tv": 0.23786875248045813 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 128, + "ce": 0.32782799216499303, + "kl": 0.09434077052836559, + "tv": 0.12276145397536038 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + "temperature": 1.0, + "temperature_fitted": false, + "training_seconds": 96.74807532411069, + "max_gpu_allocated_gb": 11.964137984 +} diff --git a/variants/events_ce_seed17/target_audit.json b/variants/events_ce_seed17/target_audit.json new file mode 100644 index 0000000000000000000000000000000000000000..af673c3c3971ef76edb29b4a5af29ed643785342 --- /dev/null +++ b/variants/events_ce_seed17/target_audit.json @@ -0,0 +1,40054 @@ +[ + { + "id": "event_v4:4aecd3d8d87778bab6faae73:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a5cbe3a96ff9859f1afbdb0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1fbbe349381af7a541cceb35:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4db212f04284ad6caa8a4724:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c1e2ae6e3efa6052c93e30c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5378d5f62c640d6e0fe36b0b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f24e884a3b8605a927a55f83:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f458277aaf7786d8fe20847e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c644d0d083026cec6c153e9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ccbf6b8f9a588cc5089fa918:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93036e72f8c9f2d9ea0f0b02:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34792163cd9a8f97c1f4603b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69fddcedb520a75892669dbc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb610da24b02cd384933ddff:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdb45bd714f3c6fff339413f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6b7a7d7738d1c2a9874105d7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b083ef1931c7071e74652dc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fe27b46da3da7e169a4927d0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:196a1da34c2b563ecb152fe6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ec30ea0adfeee4fc39690d83:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9fa06a502ac89600ed238558:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35674713480dd730da6c2556:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b247183873660302da213d06:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91f994fed8c8d4008fafd666:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:46a6e10623052a887d1c2205:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c986fc04c6321670f2dcf9d7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c32b797ced7f503623b6407:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c8f73319f7da5359e93be37:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:168f30e6158735dcfa4fbe3c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e69848dda7a61a77ec326915:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9583dbb10b63349ac8fbd042:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e4ecc9a3d8526c6c4c36d32:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:778d5649938871adf7e3fe56:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01b9d89efa545728f8e9ff91:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a221ee0f166a7fe69dce2d60:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8683abff8c9fda846ef54b5b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9d8e40c9559f1335400bb1c6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e345dd2850eaaca29fad266b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c83318f0abed62232f75b3b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ce9228ec2e6d3dcecd19718:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3156d2a5861f363f4847de3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9bfe659d2661f885e63407fb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:19cd809568aa6c69380928dd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9e9b833eedec6d2516b563f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80ba1b17757572d541430961:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e84d57de9d328c6af74168a7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aab346f5a68e8b0d9cb8efc6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c9bfa2a376a324b78b76000:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bbc9efd11d627b4cf460c754:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5836727011760fc93a5ce716:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0afdffd76f8f8d912a2a9810:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f049876cd8d7bc1c1700c97:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25dc1b4da5ac8363f6d93624:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fd21575f55b2a921f658d041:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a95f0f6eab2d1da5dd6ecabe:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa7005c6ed1f9d7125195daf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6d76c5dbc7687767aabe29ce:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:840c07180c531dee4b614733:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:496c9ca3430d9a20c6f7c687:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e3afb50a1a36f19d4ae0b24:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:930568d553de51b6a5d759c6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8069fef34727ee2a1902a848:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5245ced54ca26c5fadec7646:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a4a10738fe569740dd0f7bd5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a6fb81b728f8dfa47a7ace4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93719e4384a5461e7a7219bc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12482d3af71eb59bebc3ea10:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ebf4a770693beb8e19ebae4c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a793ea6fc8b9e5d41771c5f9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:17a051caafaf5a77999f760a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f14ba7d8996c6f860ead6fc1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c79f246342ca3183f470e4c9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcf61590f5cf4749aa1280c4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7928a43d654582bcc499f60a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:510689d0c457bfd6b4ad0a50:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47177f54b2ff163c6def1b2c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ff7f129a6f7732d7e4795e4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a807079a25b56c67ac66ccb2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b0436e119d314b5bf86c248:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdfd7b16eb8dff830b5547f8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df1da3e45292136ec837892a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:758a0dbee8da7e152982df83:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8d56339aea5d5a95a5b665f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a88a6a2f1f0e68bc15a24c34:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0147d79e60563e071a1dd396:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:086cdaf3d5ab2ea94df598dd:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c718fdb80210fd11b08e329:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab041a7207d659c5fef22d4a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96deeefaa3365cda24ea483d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b74f223bff9f3d72331b031:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ff604244fb4de2bd94f9aca:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8300f14e76d85672a8c70669:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:666ffe69714f5e4b0d8a1dd1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1674014088efcac3fac914c2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e87a568ac2f2e78b210e821f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:254f7551bbfb139006266b6e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83cd93517670fd958ecae5ee:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:634d8cd2256448b12a2ebc37:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:081c18fbe82a3702fe712cb6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:860afa1ef370b2e6f70bf40b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:182efededded37823893473f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3d7e5dd2445849704958bf60:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3094e2d4d0cb34b5bf1d209d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:901067852a870e821b821b6f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d3a0a64ff3146b2c4360e4b4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e856c6d0068ced00cf2a375:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98e210f7e13494da7390809f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29f555194a95a09dfaa16eff:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c33b69ffd0e5d993162a45a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c37e3a826fefc4b83f1afc5d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7aff4ff3d50892441a8f2162:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc2d86938d5220b4ec91b2ef:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7b7c87b3eceb318ec6b95df5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a46694e1cf315b823b32b1eb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1424aa2a069f7812473505b8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4532e9616b169212296aed5e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ef476d69e35e007ec2ed8bb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b295a533b5c8c0ec6a94727:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a76c9ad5103bb6f3d9659029:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a3c6b25ad42df48f7a7978e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a05fcaa4caa3a3115129fb2e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4dc051c0a01fa2169c8b9244:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41b15a05073f28387831b691:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0dc938f084a2494a91529d71:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee54ccca139e1cf56c89a462:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a54e3adc5cf2a695bd83f471:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36aa1f5ef31c1178a822b653:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a6f3af0186a53a52c1bb58ef:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e6c7466ac0e8dd4278a1cb8e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea5ba6ec793a1e9bd609b765:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93cadcde1301139898ef49e7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d09ef3a5bff29862f750ced9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddac902fe22413b8b46805f8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d973b63f3df4fc07d52ed79:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cae384c4a93e87b360c116de:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6cca008fadc530233719175:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:538249383f3cd0a4e1dd2e1d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0edb8696013ec30415cd3deb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:554d0d960937239f75a5af2d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:52b8b97eb0c8779ef5da3933:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2cd0d4d1a69ce0bc5b4c649f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2898792139f17a9c4384afdb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d3e082aa226d2f5e24dfe45:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0214fa91d48fa1b6a1f894dc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b62b892a97c579a76b0f53d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5203bbe1d03bd01e2ca478b8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cba64bb97c7ab78c0fe29324:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb8fe05a7b0c62029f2e6ee2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d7193659271488065d5003a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba98d4315a4cbedd4e2a4a59:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d255c6ef457e46563b4b57ba:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6af87b747f689c9b4888711:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67d1f85b1befe76def8fee51:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7d9482c148cf9b95ebc1caa:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1d3310aa8547411a90a3bb7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e6b9dfd3c07d3ffa2205fd6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1827722f96e17db2bd6fcd84:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ea73ea02edda40bc54c044c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b4920c9fd2938836826284f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e027fbed3245e5b946c91da:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01f48f85232115f71836ae92:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01ee9a28d0f9469c007b5cf5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac5276e3d4831247884eb8a5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a95af135a3716d9969ef0e21:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b2888c81393ceb3d87874a4f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62263c53b73cf9c7e1c14975:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9569d605fdc66bb6c76a935:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85c6a22076b02b74573a4a22:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:616b96b76ad67d4c9deafeb5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:00d0e1e49c64591e5f0b158b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f709122ceeccdf38ad920fb0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85a36ad246c92d82ed439aa8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a8173dea9319f4acbdcd0f8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c93ed08d1f079a56d20cb73:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89b8fc1389543cc85a9282cb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e88d9a95e134302a29ec302:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd8e11eb57380f46a4533532:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3382186e120c20e414cec7ae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:af836278eb0dae03ebe1653f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d801e9df52ed90515626e65e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2afaba5e89d1bf4e970cdb9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e46b02ee09f14a9af0ee7bc0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22b1e3da175260251826164e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7dc426e15a7fa672e29109a4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:002b433b16ecc07c398aa3a0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c752e38ffab6e0a3117b779:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e1f79395cebd5fb8251a78b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b0f6359c5359f4b9d09a0cd0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1f199a87641482c04355483:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a9c472da2ec4b16b5e252ad6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30e895a88955e8e1741c3cce:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a4a29c08ad1ef7361d6066c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9299075c97ddb34754fbf99b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61046bb8aa61abef3bf5c069:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e812c569fbd2ece72b6ebe12:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a1ca87d4133b30b1b878c04:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c6877dfd63449d3097ef947:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4a07f731d746b4a32a673d45:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4a54319b85b667f85c1f87b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3a8aa14cf117a3c0edde5594:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca30abbc0d1a82277a7df57f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bace6e281ab7070db1cc530f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9bb6c22f14d1dbd762946aa2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7c4f642e1f7ed06decd0ea0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40bdb7a9c10dff12db44a3cb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b117b65e1f5ae573455bc7a9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a54baff6329b66e6db767bf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5533620bcffa09321a5d0f83:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0b73a17456822038153a599:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0d6c54e96ac000bf11cfbd0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d13966c3efe0950ebbb60c0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40ec2c07a3cab13452482734:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e2f8cd7ab90d301bb1b973c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85bb2b2655da536beaa4df6b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7d124aa6d2ffca932231b33:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4fb18e94b72edbadd4726e49:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:199a56e347100553975c9f2f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c36fdd7d4b562bc197282912:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8691b9f59d2f19d6fa25801:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:833b703763757af5e4b68b35:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:52720ad7a21f333c077b2110:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d932e48b4fddfa51788a6355:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:252d4d472631153cef78cc8e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0168631aceec340c766a850:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b80331706205a42292cf1af9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:248fb4f0ee7f5be24f49a69b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b389bc2b0a17e007ec63866a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41b2f786c19ea8833f6f57e0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54e3862145e26d2759ba5ad2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ad5a0c8f22b27c96b4b7a111:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7b9677016d81a9a7b746276:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e78bb8f32a3125554b5eba2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d01821e72b6a2a46cf93409:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e983594b520fb7548ca6a6e5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81ad6fb442bdffac10981404:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50bc7c88108132050d595467:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:221e3b56cd290837c7f03085:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b759eabb218117d5b4b1afed:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:549f1b7f97212fa1f78a3c5b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2493d89ac0d2715ed9bfda35:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30bf9428df242ea371274770:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d8322cb4fb052fd1ff83efbf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c4fe6f890052ee219851997:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:04724d4d373009c3c1df0eab:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a3cd9d92de85f633a5dec4ce:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d8481054b13cb02eb33a1e9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e0898e9b40f457a17d9026a7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b53d5d0ac873cde248198428:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b426a808a043cb0bd822ca3f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f70a11c264b317566cd767a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:43f61961b03ca186cd8212b0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2477ca30a8a97a0e4d25416e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d914d7d82fb5354aa46d8b00:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fc045fed1f8537f357cf453:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80243610db5da348b2a7def2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcc4e6fbc8c6ed9fdf00212c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:987b74ea0bf071b910f7e810:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87a1cfef0ca559c4f66887a4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c17fc3799d7c063b5c4bfb3f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4a30c805073de74309d662d5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a41ab14b6ab104574384653:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d01cc3d9771456d0369920b5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb9c8eeff8bbce90cd863a26:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb3d2173c7b2328ec6752939:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0652e8b64a4ed5ba0d647126:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf59db09948fe96a90130fdf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c3adceac18a6d6e8c481e1b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba9a4a75a947f371d1f5a481:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9cb3ba48fc6d5384eb585f7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:292d86322c0c34e7f5c7f30e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aeefccbec51a478f2223b626:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd71354748063f8d30e80b4e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fbaf2c0c62622cc533b401b1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:239ef25f5d857b47d242d6f9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98252467d7707f75aabce2a2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4092f7aa8f3083cfde4c58a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68eb85b4f7c251c9d3263f6f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7361c28dd204b7690a52e4dd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:236c20ffd93d26e8a865b73c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f3fcb67d58c6cf28a08190b4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b2add6850aeeb55fafb5066:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e439ad1b03f92a35d30381b8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:74db8b8082d0ec7061b0dcdb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e595c3863ba7c0335d80f775:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1476189f52ae003116af8f2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:104463a9e824c39ca1c78ded:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:679deb7673d4ab6a1ba5806b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fbc6594a3fbc5d7ddf80247f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2668a50afb39f45a5708eb79:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d578a3b877bda7b37b471446:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:938ea4a3f43edfe150282098:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f694ff937dff30da47693fd3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:990d3cba4db1fc53302fec38:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc42a7f86d3a9a6eb121adab:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bfee33f042bb50631e54845:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:789715bdf6ab3a2df5417bbf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02e0e060d800c08a7feb46a0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9432f491c7a3010f7c860661:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e691b22b64708e4cad8a4a6e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9350dc6743fbf0e2f2189507:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61a94f388e9f8f2dbf14d11c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ed16481b7895086a6ae0910:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f26146bb5865ce7081895d45:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25ceb291cd0bf06ba1da25b9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e1a74bd5e20eb6c116f78865:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3550014400bc31035676c567:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6369c20a70752ff7ccd6688e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5e6366bb9d2e3f6488841909:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d89fbc2a6938a5591c102d0e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:807bd7bb545f3363c43bc83a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f5c0e4ae9fa09e328c450e8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81dc5e532c6f1e076f70eb89:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7bd683df9b5f1f554ccac45d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e97dd3e41a15710dc9dea1f8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:214b3693143e52555cf9d3f8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3f1debc5bf500f6a54ed3afc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab0b47508e943da5fc0bc22c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6b16b1e88101ab6ed3cbebbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4274505c3a6b1ea2236ea6b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c10e2d63cf9593db22f07112:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7615413c0accc18c251c8e66:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5b296bf9fbf65a3f5eb2335:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12de4df791b5835d3510a9f2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:605cbf99ca463999d7cc4493:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0df1da747fa694bcaa207045:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cbd2aeeb33bfca228cdb7f9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddb4b15cdce7f68834fc04c9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:71044f235933988434f26106:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e1b42e7bfeaae483aded5f2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1932357263b412187a3089c2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8da659044d73f614b7ea22e9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41b49028c2c8f400c9467e36:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:882efce7e9d263e6984460d6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcfd6d606d9043e8ce82e88d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d55f1182f70611ed3875206a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96f2217a5354b46068eace99:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f7fac7cd2de3945d1deb891:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:002a53a5aceda2ac2ce52bde:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1fd73bed926baca20befaaa0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c5196a2731657dd74eb325b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e85d3f46b855dad1c5f1ddb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:255697241911fb3c271cef5c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4048f2e42e8eb20762ccdd6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:138733578bc678ff32537417:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bf24a4008d41f915e1fe7f9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ceca5ba33881a0df60c9fefd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:99ca534f1179e15e7d58b376:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d68e4967f66d33dbd27f5a0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb2146f6dbbb9fd75f85953f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ab35fea550164a919ea8020:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0701685f4b125ab98b01f43e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:197a8ff15e4efdacf1d9d8d4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:570a426861edde554174aa2d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7ec7d81347630c4503285b9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:797e6d858c34074a543dccce:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b995f52aa70685be6dad342:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e22e5f2f2f18e13030c01be8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4d777f612ecf72fc257f59e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98a69fe82363f91bc871aa9f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e15a945584411fa1b32245fc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0505d0c90f7e5a0ef9db8097:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d2fdf33219471e56a1ce28a1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9cf1ec307abe4bf0da3ec2f1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a50eea7365b428382fcc3f0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68ec8b82a6b56d72d79afd9f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 289, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68964972963ceaa0251fd85c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 289, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6e2d754f6e8a30dc996b714:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 289, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:79e1eb53a8ad8b73f07e96f2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 289, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba7a4d97d9c5cd5a91633d7d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c01989785a6ed6406aba33c4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:349a5dcd9f4878e6b189e285:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50b714be855f405803e90240:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87b2593405d421d6b7a28dc4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c199d672a69e2d7636cb026:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8280d12ab7784126b42b0681:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:721054860789951044116121:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cefcae3de562b00d42c08dc4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30daabfe807e170a81e6eaaf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8905f1c83713bbda13b0616a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:358db027e477b84d72e96a03:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0822effc0b54089eacbdff44:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0df480a355860888d0a5602b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6135141e7e7b1db72833bffa:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2ae4acb563ca7fea8469e8d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0669a64365763d5c2f696236:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93db94cf414a8f414c8d615f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8dc21220b05303caded38e7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3423add5646023034abf055:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed250b8345eb1d699b876343:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11af4b1978a1f0af8421b399:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61b84dd7e63312b7d348bcb7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b43548aca5107e43226b56d5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34ee612afad1e99c2eb710ae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:108910315ccec0a3e244a7e9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac8cd90feabb6eec95839d72:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb6d76511fe15121ee9832fd:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92ff30de0217490919c52f65:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06a99381abaa185a6a9fd049:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4a247b41be9e261216e342a0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a774bf3a92f5935173d3d2ad:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30bca1ff0db749828f0cdcb6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdea874c3af71af330e8dd19:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8d874354308484a8eae3f0c3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0dc3f71b2a4699ddefe5d14:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ce616f50c50ea8f50311521:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e57fe3476a705cca6726fb0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdfa2566ffdf094b245294ef:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc273089829f9de3e5112353:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:975ab57af5710e33e2b59472:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d7dffaf2c214fe692bde8fe:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:17e3d8f8e49028635017cd03:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9fd92e91c1dff41dc113b20e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:302d27a5ea6a934d5ac42b00:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0dc4544416ed815d7aebee50:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2243504d61f3609cb6e5a783:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69c5810b4917634567135eae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92ec2c46cf996dc7f34871a5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68e607d993a088bae18435ef:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:215140a4cda26a73503d3a83:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f854dfabf546b974d504566:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea3b5cc1ea8122b10f5f86b2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1350d478ab3b32c8b9f6b2d4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:550880c731ebb9d5c09ec24d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ea00ebe76fa3029ac6c7fb7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:476cc169236574a7988d1204:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a55e8124c6fa131a565d4b0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:303d0afd60f97e56aff794a0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:14435878a8d37bbdcd2dd5d6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a5dc91b79c9a2a823af6c019:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:60fdb9d681067a1a91f0b72e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39cdec65c8851ecd19231221:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1e923c9fcc64fece7829151:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d725ba1fdce0dc5b09a394ca:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef90e306cbfce9424175126d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f6754250ce60629b2509439:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ece0427d8d7392715d70ec5c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b2901b20d25de32fd1f255a4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9563b40907302e25d13773b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:57fac8334503978da5a90a40:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:347d698cf044ca9da0e619fc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:53e1ed1fe556a8e1ca5c2c4e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01e0e74ed27a6dee36461958:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c8fd4f874d575a1de9108bf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d740012301484b159c0a0dbc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6c7874bfad93f70a1ead4e6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:030b79e3eaa3583b3ab8d5b2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d6540ee67a566ec60d881d48:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c85ba39c8672f608029f5774:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee2ce711eb4c97ce04e46bab:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8f7a4879de75203775386680:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18d809d6ec9834b79dd8a863:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:64be282f7f43715069cf8f74:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3a7ecb6565f50af24257dd6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb0db769fea9db7eeb4fee99:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:710df1cd8f2b3a69905f05b8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f370e94562544c6ef17f2dbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a00e5d8f9b4a206c81833d8f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3356a702168cbcdc97cda7bc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12e939a19ea31dc542251c43:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c536e43ba623c319d126fff:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:14c6f58596354ecb87920614:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2103a5c5f4a45b95e0d6194b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4f9de9d504d227b5a994dcc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02e2cde5d8ed423d57bd5ac2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:908b506f0688ec79e57590e3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38994ff5042b36b0c1c55692:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b0f0c71ffd2edf8da980852c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:046c8c58d982cc8d5a0492c4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4e0166391c4975325e16209:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85a6940d1acda164739ea4a6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:56eb58aaba3e308c1417b500:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c1ce6f064ad5b498c644532:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d68e01be9b4c7ee47e12ea8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba040a21090fb09f79e69665:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9806fe49684837640b12d62c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c1ddd083656018f58c197ae:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5c1e390d768827716f5df84:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:79649dbab1f24cf15a3ec4b8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30230f7ea8af16bb21d90e88:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1831705374f85e0a6e15efd4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5dc9447d071b881704cb146a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:118905145afd251a26d29e6d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6e96af1995885c22eeabc35:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f478c7ffc40d4f4a0f8a2a61:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4fafa439945d3eeb7c614c61:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:796e4152d808101e800aedaf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3610a9e1816e92532cc72c29:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e1d84d5ed67202e645eb63f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89c2f4199a905249e1082801:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:55febbbe1c3cf90506d136ee:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bc9947dd9ce76fa7a77e0671:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba3a43cad6fd0c928b4ea445:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7356066daa7cc23f0e0badac:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b575f801c426fbe0ab34b1e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d7036fb11be505f0a9b3a417:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6923d5f85c7161e63d26817:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8991a01dac500c3e224d55f1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e2b4f6fe5f3a3c9d0f7880d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cb61a080e4ce926cf069537:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6193292593bc19879ab8d578:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c234b0c1046e975e8fc5e09:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef728f9a49ee671e899af858:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e24f353136e07d3be10cb3f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91f7cf44076c6002bbe8bf3d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83548f2cc37415c0eff38d44:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:109ea27a31551beee79f4482:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e504926c21cdc72076f9db5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2359ed68991dc93c9796f8b2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:72173ff70c74944a07b4ce73:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a7c9684b7720da082c99716:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bb0aec73c3ebeffe25c7e9e9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:110e7c0c1c03f20e82096244:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fff52480c3aa521aeb29bf51:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91f323f74c229b1247c3b9b9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d1f435cd245a901f0684e31:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:56f96a9f41e776cbef2774c1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29ff8183f5bb30345cab4390:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68ae6bf18b181373f8d6d4e1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0c3f4339abf8944e23e5215:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:717557dd783541f595805baf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee59c6ae58afff53499628df:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83da5cf3d1b1ddf8fa763b5b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5c92a5701eb6668bc1f8fa9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e469a62013803fdf94f1f59:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c2a82df1e5467aa7dc6660ba:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:26687c4ad73cc5184df26a15:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fdb332d5f8e4a0d8f459118:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0990419723e11fcba0affcd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca629ed704a66b1357bac779:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcdf7e375cf726310e726038:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5399697cd98f72753ff4184:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ecefb3605c03f41a999129e3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:845e97f19a29261cce2314ed:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29f234ad5aeef176f1e44927:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aea75938c86c8105d0feda0f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbec129f76b8c83c86381f95:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b348cadf48a3d03624ea0ba5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4035835d41d16c3a1fa8b1e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fece30cc4adc4871484d458e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fe00df43f4fc0eb779df708c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02aa23c6f283014c1abe9676:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0fe3e35ecdb9a7802759f3f9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:32d2a54b10ddd91944c573a0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a662e8b20507e285a191bdc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8ab283396653bec8d845a1bd:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c472a5c6051d07553cfec417:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47b61fd930b156045842a577:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa5b804e708259e0c04ba99a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:90c51dbb35adcc95e4ab55df:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ba46f215c4d34e0853b6ded:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b2be4263a3dc775e51b23c0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5427484a4262ff3c7fbf9f79:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5d3f0b661708700e8809d61:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ac884599d6dff56fa18f63b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e696748406acd478daad6de5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f7ff868b5c88c0cc528c43e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3514d35e6a4c8c67dabd6266:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ae1c08f689bd686984af0b7f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23fbaf66af2fb9a5c43ac00c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cba82a24afb9a6c6230917da:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:42d501bd5d2639615f1bb65d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1a5c14dc584cdc571cb6207:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8f4bffd9564951db12676035:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f1785f4091164d7d3300e26:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:420d69c51d42af2700fdb212:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b46c387df42006311cd59d82:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bac467952578b22f282d77b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac815c6135999e5968b9143a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bbae94c54a2adedee667f34:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e195100b5bf4cdf9f9dce4d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f58ae09b633370a388beca57:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23e073489564cbc08d4bacff:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a9fa261051646e4c425f6bb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c2ec9a05e0056357c4a8eedb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f29c436c6df6cad38deb3b9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:09019c96afc17c3bf63c15f9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f667b9249bd4af8ee7f52b5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba6d482786393b7dfe37c6cf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2be7856c2e3c093585329281:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1fa20e1404323924dd8b65a7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92a812459196dfd57fd7c593:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d2145a80946516e84732608:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3313bc5d19b525468ec7bfbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cec6d95e463244ba34935e60:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16ae671da86347a735e04d27:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15d0937b9eb7bb64031ec594:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:48a65de2cc08cfe9b53b76c3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb634b1bb36bad706a7084b7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:babc582a0929c2d7c07c1c17:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc30f2c0273cbc1bb0be8fdd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0bb4b76303c187ae6a16176e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e4375c60569dfb66fb42573:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:af85993cc9a07a6e132ae3c5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e328f9c46b379e4a9b37d08:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e7e58cbf8db050114101723:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:177265c717257ef23e3d69f1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4250b698d7b9c10b232469b7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9dc149656015f6baf4f46219:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a726bb28414368e00ea7c716:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa1457e17b5514610771b43c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06652b5bc00184010dcb89ae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ac6b2b50471d709b003454d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b2293aeb8e3eaaeb1d16d0b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b0ac2acbf820ac72b0cd6487:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb37e69ed7dcdc938d2d52a5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67d5b01b97a26c9573f4d52a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a04687efc794322704ed778f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66bcbb7e600cdcaf38db10d8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3593695e6fa4bf47bf7eb60c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54dedf522236b05f475bf13a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:792aa4e64f9a2e723059bc0c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0ddf3066fe7a07c046f2376:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a3ed1939a8edca581359f60a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:27327bad4c5f81287d5139cb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f8cbd20dcf4f9ddc5c51d646:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:607aefb0c7d22d7bbcf0ac33:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0fb010b04e5ab00b94f9c722:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cb5cdd9c1b9dd9fcaa8e0c1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8857c7cc49ca8c72df0dc226:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f2dec7dadfa71e1ef0479e66:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0acc8b69146025cd4934c70:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:57f0093b52ec2b56430a80f6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0593b127ed99d3616c1584ee:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4cb5f335687cd8b201f6d4a6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:97ebe26d7497d43e821b8e28:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9cee84ca5c088d1b37392d15:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c4d22aadb729cb8b826be92:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd787b306f5b5f04cab14c66:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:94688c378413f969e8701a8f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8dc99806488ab6384e7aabce:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a3bb296ed9c28ff8b9a5d743:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2af1038af5283ce6600f3d2a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5738c688f78baf9bc6dddcde:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8f9796158617bd8c568a93d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4355fc17714a16342ec2b97:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9af375fccf968cba465ad0f8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:86dee5796e3794236a4ab2fa:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4cb9a623c1921ae41eb6e0d0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6b8c3864a7d66bf960feb6f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96e2c9083ebd21ea07e08c1c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b1ec99a0da91ebd52a4bc72:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:97d1a8e484488d2a6ed47ef9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c73dfff2e86789c2afb78b9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5747d09e72c2bdd2c192de3e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b219c24196b6b47e8990a71:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ce736a0e9a327a1e3ab28171:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df3045a4d9eb07f99fca7e99:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75c32f6c1b3f14c4bdca3a67:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0090f026f9e8985f16cef679:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:874fffbc6e8b724a715a1443:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7b8ec24d53082094ba6cfb7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2cb5e70aaa43cffa82c7e399:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c17e9d4a789e6976eb1205ed:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:17a10837f8c166221f9234fe:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f7df1957ce58aa1af8ec6295:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66f14148c22c4c918d32a583:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:068bcb7a7d9b80b0211ea29d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4567f0f02ac21971af6c2067:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5970165308c46d21f689248:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b820c12089d45b425a2f143f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8db344327e8832b431640f29:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30e5aacf3967d75bce836157:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2aacbffc4db777764d114fae:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9791481a2c8586a8b9359f89:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:150e2cd536a5bc1059f65dc1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:71034de81675fc16e59ca900:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:432c9d646994c0bc4a45c30a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9fc0ffc853df57b1434b3f93:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1cec312461e179bfe87fc6e9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc228262685c9af863e04f40:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0bca2d66187e300d825b0e2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:185cc5fb91e6085cf42bf0fc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ebbf8f64e89c61e0f31df873:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:256b930168cae56bd6137a15:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7ab006da8bd3d6f8f094feb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1dc1fb1bb2ab20d549e1b067:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:801cf4e3380a9253f7340871:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c8179a22901be601d02a0bfb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d22c006c056739961a9fcc4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ba08d18374086eb044a9a59:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1ae888b75b45fc321ae31b8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e182b61b80cdcd82a8afbcf3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f61c18f607c68048efc76767:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:760691b4a6b553e8e6753666:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5bc6da415ac3e3139c77604f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4f74c2cd88ba1556d014cc3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b362adce01db8b4de0df49f2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d91ef6762a4834e41e226d7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:da4878ece31bb24a97167d65:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f90237ebe0a2a0b99bda2998:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cbed873db6aab5eeace3694c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d56ef7f829097942226c38d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d56a2f30696951e16e25a68:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22d3b43e71ab03e8c809e636:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ce6114047f6b01f950d2c7c8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c68b5115e21bbffda6349ba:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d75093dc65ef08846569126a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:af9403a01d459574d8e44d99:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:204c301980a3c082bedfe1b4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96bedbaf1dc4a9c585bcfdec:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8cea37a22ec1d8811e1316ce:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e234ee54eaf6e14a7a3f023:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf8cac29d1b7416ff9a898b0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2fa298c5736a119c0f39ff7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4df98939ec82fbabd65766f7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30c9c386bd32f1dad5ba2fbe:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18eb3dba14ceb391c588078c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fae9510eab9a0c6118233776:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a45132adb7911e7510ad744:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c0f458485c6a396c24f942f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:20b549d9a6af728fe092ca68:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f9b6fa2af2d327776b882ae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b057646709e1db1debfd35d8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69cc59bd8e24338a96430932:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2307ad191e5cf6f9c229a277:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a08b246a5ddaddabb4b86f4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:799ded37832184f9b0ee3168:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcb0dd55df41f2ba318a1bb3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b952616a0a1f108d33bd9eb7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23e2df6c0d3866801ecb2c45:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3dcb436a16a6492943a2fe42:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f1af8e03283abc44dee7859:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:046c78af4f31b3e1ab065ce7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:708587746543a9710eac6c48:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d22aa8347ce5fb1a6dfabfd4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a7fa96d2e7cadf4a5f3b6b5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6637b24f2a43ddb345c43f71:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80890ddbc902f2364a603ec2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:60fcb58976b841328ff5a9bf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:229a9b464754862682004324:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddf5a73e9542bb34db9abbf3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6278dbc5ca5e99ab00b0144:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a300c51301010ff18ce3cb0f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bb87e2d44e695ba430954bbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5b9410e7bc5fcfd47057e921:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:185eb6d6fef5f701433b1220:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2dacec95a2422de285a8dbe7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3d31488468fcf83a2c803e4c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:76cd3dadfb8a076af5a66d49:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5ba5569d43760dc324e2833a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b714492a3f71948f1882a11:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92cf5eaf90736c9e783955ff:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62e0f1370c4c1b99cdc6e5fb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:26b949ebe9b83c3aa170003e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66d868a7bc10477729ef2a16:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 595, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:618fae9872bb340930b47b13:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 595, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c7cf7b900b8f387ab2ba4ca:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 595, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7566f28d2f41ae8fda8908da:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 595, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:049d0bc4ffad5871cc6c21e5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e85148c1ca8064cb491e1ac:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed45a43e07ea52f3474f3c6d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa72e2e3de57338165467204:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:496aebb810c5df7c3fb73488:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e32308b867c8e6087e133433:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e984765547431adc0fe178cc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e4274c9d2832ec90c5d381a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3389c4d6994e16bac4135b4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 597, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef0fb44a5c127f5022120f02:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 597, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1bd52436398b2ea1bdeb6a0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 597, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b30fd38a909a162fa7d896bb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 597, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0dc053390957a7149eb3d08:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3b41c608f1490ca652273ec:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8dbd0e73903b75845c88813c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddc48f37fc6e4f36738bc86e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c9592fcc0e2bd4b29222061:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a57cff3bef65fd2e9fbf58c3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c2f66b1dde9743bef20b0266:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:337b5e5fdcc4fab7d3c53ae5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3f605aed9542453a8aa3b21a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b83faf6baf4766897e6521ee:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df78fdf4f7e75de43738ad2e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f166af9dfb76b10ebf7f879b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a78ff1ab2f9b43ecfc7ff3b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ce7589b9818b5720cf2933c8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83345349bb580680ce39a90f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5c5d234c0bf971c21b140ac:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab4e71d67dba088cb00b6bc8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 545, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:200a325d944d71b733e57056:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 545, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a7a42b1c7fec28886a15d87:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 545, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9e494705f31d7673ac4f282:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 545, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:739fe045f64a353251f53255:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb79fae11d1e445c1a56ef56:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d7fbc6d82cc3b6338a9425a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1271717bd6c4ad47ca5ab540:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd275452a7b4561d073c7d2f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85b9ff553f1ee51eee4b2ee0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:95e6d29eea890206b64a2b9e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a141648fd4d5e30c21cfe742:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6bc574010c1c62fdb5c06ab5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0bf41df877960a294bd2dc3b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:988af17e89f3dbd0c2a6d8c9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:996dc372ff31a1dbe5f80051:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:091c3f8e4c36f93849ac9785:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:079f9ee4be2ddb457026a050:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8bb5f7b4bbe51dc82621e3c4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:17b0baa120987bb826e6293d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0299d1ecf01b99211039b2c8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3a5c95f52cb7249931ae8054:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df8c537529a9f98e6d25c493:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d4eae14009ac454c6dd3050:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:062eac1ba14e29549c90605f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:782a9ce4f823249e70efd6c2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b1501a6246bc51a2aa4dd4d7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d264e421eb3650ec6c88f72:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d292e83ed35deb88990c23a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b51485b0f6cecff843c3a3cf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a64958ccbaac405f2649562:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2bbca54f9fa3526037a45d5f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be9545bc7bad816745095a84:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d971b583c5f6549dda4a104b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0c431c169d5350813868fcc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c463c601d22ed6f862262e2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d38c8882612470ce3b7dab3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1017887fed6b638c2029360a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:669b7a1ee069c220b8bdb64a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:73c1d9c12699beff2c57db5c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:604f087db4066c54036e358b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:518fe21c787ca08f888b605b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eaf3895d8703c0acec712ad9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba8ce6c41c9ecde628f8f085:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:060e5d024b462f4ae3b4f91f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 571, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf7e6ec22c3105e56dc1b43d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 571, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:97f6f5a77d290fa5591db5e1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 571, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c211443d389ac0b2eaf2e0f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 571, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d617576e2f876effec9bb0a2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca9adaf9abb1b02a390ba105:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b855b583ca6b1e9350d9aa73:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18833593ad322bc2dcbea547:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a969259678ed90e865c28fbe:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b4535265f0a2e35d06e4f570:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb48a578a464a1251337a945:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50948d9010b0cf9be4705043:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c6993e692c8ea421debb519:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c71a6020390573b60dc5648:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34e76f4638ada5d0d6ac7788:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cfb6b092cc0c84712b1aaf61:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3fd046e082b53d5371704ac:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:edd3c2b78bd679cff92b51ef:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e23870d7b4eda111ef3a4af:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:73b4b18f273a7d95d9e765ad:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c79eb4114c40e995f3c39e21:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65a2de6b84eca499474c8690:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:42b1e81510ad3aa3c14b7cff:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36c0a1b384208e83ada34096:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:977b3d8d36ad484117f35b86:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fc20dd3298a3c6bd8fcaf067:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3dc84da1587d06906545f61c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7cb0ac616386bc05c529e74d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8848bebe5d1b3ba8cfd9e0cc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6cc84949cb57ffcc2e00fdae:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8aea77496507afce37ecb5a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf5c447bb2d53e2bc6f8be9b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c405410739966da6329ca882:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 574, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4590b4784259a94ec3af0811:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 574, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a1837d609911a827f872b9e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 574, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea3dfdae3aef291e7700e753:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 574, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc3ae0450bc3fc8e68368bd1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b396bfa34a9d31d23deefa3d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7cbbd588f069c8555e06a379:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3237771955dd6b30d41e8c39:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6df4b63890e6a372f2df7e9d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c152aa8999fa73ff99ef885:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b025645c9360bf3a899387f5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f34cc937b036b479071a9d0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5acbbde21967398410a1ad3a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb9dc3ab77485e3a990e3d56:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bc767500ce31f92bebd87dc4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1616008ece3befbcf7ac3456:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:984309771153005679f20451:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba71bfed0a26b0af1bdd2695:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ccedacf25cac3995814020c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b126e98cf18d82316644705:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01718a17f99bed558550af79:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0fd7b0b208e08b9f3344eb52:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22a3b10970c4237e0c8ed51d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c5a6ad773b9dac3e0651374:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2d169c2d770eafd56fd4758:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 470, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3a59d88fc557dc149634d73:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 470, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdf26d139fbebaaa044c2fa5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 470, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:055c56e08b4932985b34a473:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 470, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96571654b9ff81c52be288fa:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 471, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7196e90067afe6dd404174a6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 471, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:13b7758b6243cac5f4ce4f33:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 471, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29e879f91e5f944ae6fb86e9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 471, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:72d4a5c83ba0dc1ad7b77ffd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa1fa26103d5faf7c72e5608:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:074d4493dacd1dafc762a069:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e63620ad96cb32f54bf0a456:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b139a1805c6830a6e0346a13:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7028e58442b3c1ca5f7a2bab:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b48536a855b27c38c2b91f7b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1f1a3eb1f7ddbbdc865bcaf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bb7af29c3d9edb7c43eb5892:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4a4d7adebf309811c1e26a6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5d627257bd76f3bb531e065:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c8678893c0a76919aa83b1d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07501ccb8a07f49088214544:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e464a3a08ac314196aac74aa:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0027d23e3c9a1948f746974:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85b5063a4d5883992e6d8acb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f10fad0488934eb0b7414f32:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a56230a85cabeb26d046937d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7ea98cdff97172aa7909b3e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c6dbf140264ea67b36f0e22:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f253fa8955339167f5ef9e7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c63bbee2ca96de372c41de0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fbd1a1bae32dde189574541:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1be4d49642934fe08545764f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67f61a4add1339e5e069b739:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c346f8cacbacadb164856d6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bbcf2d453d8ff3bed96e9911:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e656e8fea427f33118c16387:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3f85c9e8efc59b4fcbbabcad:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ab01cf624903a190a6f9c45:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f5f538db7a194b4a2dd3e12:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6db135cb082032ee533b729d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b572035415c96f4ee0b9204d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7fd2cc20c2d4c0b6d47f2e1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0b994dc9230cf4cda2cb02e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50db9e62d29594e9b708142d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcf8541ce0550e598a5ba3dc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab8ae672e41046c1e9950315:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61a57e79294863efbaf08741:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e56b912c71d445003bfb4a5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:483889dfdd43b2b4b67d1f0b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:88c6ad09b0de5de5a96dc1a3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:618baa94527b6a8d7be08d3a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c84d1ea32238fa2cf7e9f3cf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d44348b24ab9996c8fa75bd8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85f2721c164d94b1de3aae76:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c3a4ccc6de6097845cb1a04:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a5ebb92e1491f02dd8513c5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b8dbbc2664243f529d57120:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5c96432e55ee99de9fb2b51:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:88f5725c37f73c24b68212cc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c26656968a0f61cf1c796605:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33faabbbfc6433ff4d689a22:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8bed7ccc4601dbf65880a3e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5cfb3ca0b69efdd6ffa3a4b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07307ec2d55120b44f4195f0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:287b4c2be4ce7f890ec460d4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d19e27c3ab06efbaca54ed8b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c21c6ed160ba9b2d5d46b2d9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4987475bb3c5a76a9cbecaf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ff71f7ca6922896cf864dbe:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:545469addf9c3e20dc5ed800:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3da02d11ab3427b378923240:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12097124a965cffcf6e69a6e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b4c70a5822dcde19ef83d72:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0fcd4787a5d6c923bccfd47c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4dbf1a5fdadeb368ee2d6fdf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cffe2b5d1d8fae8977315f86:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:615fa4d6aab97aee94c4e8fc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:147608a594478104c0d680e1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:45a81ded48065fad628676f3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8a1329ffe89c64471af3ff8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8fd926445bb1a824d991482c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ae74680c924ec9e11b897dec:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a538341eb215abf1380900a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc6390a98c2cf33ca8901588:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:341e910793242946a516b5d3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a451b75cbf44c73bf74336d4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab67bad44e9e5ecba285674c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1340a3c59769cdc31710736:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:600c4ef94cea891dc9f7fdb1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:959379badbee938823522893:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a958830f5b4e68527d62976:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f755b9e0473b052c0a3aa602:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b31a1d5a3ca6c179e0a8838:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dfe36dbe4acef0776455fc99:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40b79952844d2f3c619c1345:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fabe6d0c3e9a7212b5f81915:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f7598b99e691e09e3a1c0b0b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:94c7291587c7fdf8b9cd1f7f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d66f45b2ff8b346064a296e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e3d42be8eba9d63361c893f9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5ad7bbbe7642013833ec7e1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5908e06d34bf003d7320046f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02a91adfcaedc94cd3fc583a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01ad27da1a36c9cfcfeee4ed:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75e80b41dac1e7e7e6338d60:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d89f7f3845a780149a83083:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6fd62684837c029a58c01961:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40070ccc7f7ea6b00c53822b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a118cd61fb57f5213613afac:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47f135bd19c3e53e991d8ae1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89a6d774cfc7356312c86566:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c689bc24e86a52674e7a022d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7fa60af1f417869daa465c4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07f166996e83a5c86e7d0708:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c721d20f39e8a865b3e38dbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a17f849dd131552ee8b1c0f7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c903d96368e8adb351b9fad:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f3b1fd80f7ecdb5a513b560:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7bc3aafe23a827ce7d229340:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c191e4c033dc516e2def8b7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f9b23c73cb7510a5a2adc3c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3fad815b9daffbe49eecf24f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:48a166fed1e0248e0dddab44:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:63f6cac5125a6e46b2e48347:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91e20cf6a48504b4ac254a01:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:519b1d4725a45bb41806f62e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c21cc954aaf7abf7c59f3b7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ea4788f85caaadb0bead233:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:493ec62203276581a79afaa6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9837b7c810380c92f975e234:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3d525bb886f99fa5d17bfbac:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25d4490a9bce7ef88aa34685:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34db7cac091b269184b7468b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:24ec67ac5fd9e41a406e36bf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7c85a52d4d1ee808facc872:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:32f8d457f532d2dfa2906b7b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:03c379704bf9c1ebaef44f82:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 468, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:781e96970e5b9263f95e9471:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 468, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5cf07cea038f1538c06dfbc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 468, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b23126dbe3242eefd1712c3d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 468, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3212d633a8ddf2624acaa255:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3ccd8cc0a0446a25bd29e026:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c642da1c522c16fb79934626:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6780f90bd4ef0ab4c964c852:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7ce50cc4380b8a55d34f2ad:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4b222779e4ff9d85881cab3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2328e1873f3677593b45fe70:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3f8d269946cdd77250c0ca6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5bdf970cd0ba998b778b5677:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e33ab212ca290dda2901f5f1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a45b2db533539d94499ac60e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f50ed6b0f102353512c2c0f8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8e59a6e593d8418cb6710d4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22af5bbcd4d95a783b6dcc39:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:145fbd6ede95ce99e4bf7b3b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e2e801e11135fd6f8613f59e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c3802c59affcaaade9387bd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:945264ffb4ce0fa9ba9f62d9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:13ceb9c9e5fed7cd34292306:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c162b5b08fb13c438da4fcb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38df9ce8f0d3ae77cd5b3d34:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 557, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3d3d83e371282b03a9d672e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 557, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3387f9c50b8eb2eb5bd70c80:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 557, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3b888f524c2b32f762e2701:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 557, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b20e8aff695f7d24e9be2b28:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9bcb34864cc835f2b791e60a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d028051d64f42217987da93c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87ea8523e8c98555a7446120:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:99abb8844b2fc5468bddb7d7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87c347f39283b0d43c77b36b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:485e1836d263485f014bf040:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f8688508369fe71529e15548:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:462244b37ace954a228fb0e6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75e7bc080a392a087b2314f0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b43153d096085d3ad116033:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1aa4fde82e480d4ac83dc4c2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0da2d71ea98b488d4825c359:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a054761dccb525a4701407c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:591f3eb85d17de8b1ac7f19d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ecaaa213a8ac9a46b6270ad3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cbfe4b454560682d4dac3674:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3d0301825674c044655bf054:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31d62825e4e3faabd7cc068d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e981d07ed9b3bbbf0981a372:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b1ab97aefb5a673e3dc803e6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36772881471abb9fd040e64f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ad44eec89c2a9eb1f3f8eab4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7fdee6f4441672fc1c465c9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:43364de707aa5df6deee5816:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1999401b40125ac07cad81b7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca0dcd12281927da367a76f8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dda5d42b55829593d7fe9b0e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:727939aa5a329baf68d7a0e0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f95b5688118480526695ea5e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc66325c50583ee4fdc23b20:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:59e119ff87abe2b42127574f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6b24c14151cba9f11c6e787:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:94a76bed42f3b0e80089d539:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f9fa00b0697ea252e9a99625:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd40fb140166100c77265ef3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:74a3b1206e2a1eab5a52bf1f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cbaf99e91292389c5923b2fb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54a6e2e10c2f2cfc6aa6b776:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:afd58a1a0b150557c4f358b9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61fe10a39e7ab28f99a50e4a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f52f2584a7c72fc9442beaf3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d914395c0130d80e25f5b1e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9adc6a9552d422d6d88b93e7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22b5003ece8ebd2aa85dfb18:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:133375e6d8b9d78fde7fe0a9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f55652b7259c7ee47df3119:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:163467daec81d95ca3d8af23:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ad54d23d660eb5e629f7f7c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0af79b5ff2ab389db6a9b35a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8388537ee2384b30c2c6feb1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdc7364b4884b8836a3504d6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:833f6de9d608f30dab171f6e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c7c2daa5474d5a1a8bf11fb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:905af40eb3eb9c4955566266:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f18fb2abcbc23e0b64f52044:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61e1497ee9e585db875e4537:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:72db2db4d3ecbb56783e14cc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7de5151d93cb1052dc0fa91c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9aafc04b8b07637d4e6752b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c04098380a08cf3c5fd7dc4b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:839128e9da684bba271045b8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0cb40dc7be962f868cd0d15:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ff3ed1c28428f72ce3d54969:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30fd05c79c32add90cccda8a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a80205837474bb0fd10f296:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9dec1ed34053d646cd07921:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa50726b01107030094d50a3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:885ad0ef88ab51749e043fb1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f33b378decbed1eafe240310:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8efc83a4ba1f7cd1b0a3d137:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5324a7ccbb9088a70298bab3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:21bc249588c2781841c7149f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5960e3b44b02676823bc81b3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f886e82fd92bf0b8349c0131:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d595b4af79d59ee63a488656:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2908101dac2781f123f0ef68:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:628055a68b19320e37b5b9e1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ef5a13f24bb691a08eb4894:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef5aa99df625925ff164df73:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2fb5cc2e2e671c63ecea93cd:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c75bd0440912fbec035278e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c6963d8eeffcf39a12cdaf8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:42c14eb9a1972ccc06eeb024:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f398ce00bd97fc5300437285:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c95a318ca2c09aab72f0b60c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7af16f45221c1b25d43fdc5:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b34dbf346315bea049b3fb3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96b9aa6cd719032a9c18257c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4ec420319578ef4ec44166b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50090cc679034c7daa5d5dff:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aeb3a33c603129f6ad48990c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4003b845d1dba349c5b5de1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3cf63a5a4dc396ebd6a1e86d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5dee3df69aa1c8dc59132a45:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:248b93ba85c9cbfb12c43dd7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33f56e9af08bb8c54919f196:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5de3d14d983a2614a765b5bd:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:46e83cc9f0b107b1d4b5dba2:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d8237b555d94e916dbdc00b0:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5663e783bb398ea72082df3:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c223dfab341eb69c79616729:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:46404f67802b9d293994222a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3fac0b725736aa037a284eab:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16e5d6ce016de554cdbd90a9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf1d97efbf2d8584141f71ed:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8951b6f604ec215762575314:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d052df36863e4df125f517e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ee2cd16ec08fe398bfa2a5f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c2385f928b4ed908721b40af:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4e3db884b0dfade44dd9330:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8409b7e8224ab1eb841b57b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e31cc834f61ecd9cb5c3e740:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f1dd8cb4fe6ab50fd4b3e11:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c02af18406fff8411250e95:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0979efce03ce9419d8994a1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed40ee9cca77a8be8b9fbbab:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5307030384d5fe47ab60330a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:707535ae8ed979160e6b04c3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb4c68dac851f72cb8fabc88:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0952ec78fc2d73f0df518fa:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9291fe8ba094d83b4573830e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c87e0c0516ac7f4fa71e77f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ff4cc1040f6229f89180a4a4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6b873ee7e0f5fdb1abf77ee:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41c2f6007e0e8a0bdb0e0f71:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3146e1cce7399b8e28959db7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58c6d4dd95cab700b72b193c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:576db2fc3e9ac93770de6483:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f1b104a5bf261b57ee60430:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d761557c925635868a3fc76b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3191d9616353be78e31fcb2e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bd78639e23fd2263538d1c0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b9d7b74fb4d78591c2e65eb:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b9b1854124a8236ba783096:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54d3d040008e83ee425b1f6f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5a17f6e42901b9f1d1d3c8f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:224571a455d751118a1592a0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f0acbc9db544f0da7b03b9e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81e6efa00f081652c8b096a4:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96c11e9b0f4d5aef92569942:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:21935d79c6eb68227a4ad91a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f6937600ff421388af21a2e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:534fb0d488b7d85d398fbfa9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f0211ce19200b833f04fb7b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:310e0456c80d12247bf7d61f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4be135895b1292470b5613c7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:591c3cc1abd7e0baf57b6f56:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2dcbcfa3e1bf17697a7dc728:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5e23df8b5300f02cf83baa0b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df42f6342483435ef082cc08:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b581c4b5dd4a47ca358c58d0:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:00ad4502930c218eeaf66300:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:24598b2e1692167beb453b93:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6410bcc195012defc1780a3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f385b6f4f2ed1bb9687b6a26:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca558d1efedc7bc4c6d76a4c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5ac8020fe947017cb8dcc098:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:328a2a0aa9585e8b9d115453:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca5706b1e59c369c22972456:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:552728dca0751a5d929d866d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11f668887cb34681226e4eaf:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0276d049b418865881211dcd:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a5b02ca0a3827dd7e41cd7e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16ebb97c6418bf91877ddf82:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5a783eb6199ad75c4620232:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:429f555c250f7077556c6947:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bcd61b8ae841a07d4910f7f5:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e15fd669208dc30921c03369:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5455096fcc40f6e589bff2e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dd2ee3f2acfb76ab2546aa25:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ed526d0bc4f86b611cdb278:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f837f5e790c0ce258b9b477:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1bae6d5dfa00dea4a46ba23:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2538ecd39313177c5433cfad:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8423b15e5bd1044f29639afd:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e99b6575f88dec8d6c09f7d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ab549f45e7c611807b0c342:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31d8e8482febb32e62cf53a7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:49d20519fe9b492fd911b0eb:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6b15142fd829ac7e4583a9f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e768d2f79a8fe07650bdab4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d15a3dd3243ec9d39477ddf4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:de52fc939058cb7b8278b22b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6caf1df3ade6479c87147788:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:09541d962149d706515b7aa0:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16121dc0ed0ce041dc78f926:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:973d0d65ad51559ea03921b8:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87af4db526b7f881a4d5a315:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9d7a93f9a57dc7c27dd75e2e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:71c2e23c877efb17dc724b21:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01cbe0c184b8d989bdd66b8e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fd0c2983daf59eb4fb0300af:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fac77920146c5c8432abd7dd:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b54516c2ee133fb35b522bcb:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:675a9320f80d4b2f9fa53686:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:42a37ea561c0be6bc3533125:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89e0ba188ef8e6704a6d64dc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9cfdc0ced617220922eb71f6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39bd5569447e539c9e74aae2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a3df479a9e5186da2dddf934:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:19fa6556e3e5ffee61730514:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9d9a290f5daec2408097133:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8fe29088a567384e2e6bbea5:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5ea752773b4597b216c4798:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:952632a214ace986f3876d88:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:53d851e214605104686da4ae:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:436c2971fece8e8a1cd07b6c:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0820228a7229f63e299934c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:477d8ac01585dc486465ff0a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:306db963fb6d2627de3b7867:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7909d1e5faac480e7837c3f5:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fc59f5f020c78720bf8e749a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 260, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91efcc5af8dab70c42865e7b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 260, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c7c04444f50e40a1fe9bd0d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 260, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:551d132b7fe1957f1fafcd9c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 260, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f7e877788d4f0483fb831b4a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e1102b0e8c86b1205eb42ee1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30c32511b5b558094e40f4be:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ede71c7b809f014437af8af9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b241cca34679e508f80d520:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9bda0666e6bfda7d18f928e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9b387744271fb301791a5ca:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:88d44fd1fcb501ac14a4ca3a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b78e91b08f7c33555a2be8e6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:337befead30095bfd00497f6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:331a356bf87970533c9d0725:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:497348247c579141e0dbb7cd:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5f4b5af955c3668ce46b278:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e9fb0462fdde24b92c4ae16:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:397953eb4614419d1829c2be:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e1354f3cfd523ca8a9ad26e6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ae476711536eeb7a36e22892:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dfcfc043a7fa144ae814b82c:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:679f14fc3de034a6be2d0bd0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:087bc9b0d1dcdf437027e20f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1e23bc5db13a97e887d68e0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:73e95c4d2684383be003f37e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87d722451f1b7193edb0faad:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:721ebd866cffbc7ef5060c48:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd75add80aa1f487583f4872:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c32bba374bdb7e14985c345b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a021259694317bdba5827742:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df6a5dd9c406bb8ebe3d304d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:555a65463db746a00caa9303:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a13771e185ce550ecceb9495:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ee0e3af1db8eab565e6c4cb:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22fd55f789f1a68e66f445d9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:045fc472036e54cbb8252ae4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5ce09907c4bcffd15ffecda:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7f890a85ef1d61f9596fbf9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b44a56334a863cee1f199a4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:14974047485ab7c0046a98e3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a3b8ad471b72cb36bb601cc:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a89ad1dacc449a52d1ae6fba:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f89b9397e2f164a290b63b22:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c2006b576adbbdec1f51a99:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4265c05b34d7c35bd610aa2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c0690fd151f35e027f00ce7:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:43d345259f5826a942fdd763:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:305cbb6637f27234894cffad:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0aea7e42d966b3948d9e5bc7:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b67297d4c55629a3dfe7de3b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b055d9c44b56b5c3f4e70009:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29f944e58ecb95c0cab6e537:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35893d776cf8a9befbf40035:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a9c32c7625e434330e609edc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6128e861a918690c6e5de3b8:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69f0be365acf463400f838f5:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62d7d2ddaa30a852b22711d8:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e5252df8fd788ab6798734a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc1cc99f8445a13669f31a12:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a40e83beae3627e8be7e5e0:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:993de02d92df7259e27471dc:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11dd0449348645cb6381950f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2797e843c26bf42ebef56796:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b7c46a055482f10e1f7fc17:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a7264f7039f6beae0e9eb36b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:059b9e89e0f9a39e9ca07caa:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ccb8e1870088eeb3da0c8fde:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3cf32b9aab552f1682389315:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0324557b4afae51b6c3b6520:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:562d5fe09245bcb8558bc151:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d70aab02a83c9ca5067874b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e02e43a0c3c80a1601cfcac2:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e3f53db64493301d2eb03fc:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c68aab7c568da7cfe676b831:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0373c11329369971628bed04:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5931355aa7810ea88f612ff:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c8e6f7b20b98b0c3046f1e6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:44e50a716d69e952a59fe921:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54cbbc305b7176514941a66b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b89665d9169a60779097a93:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ed998e08fde27f01c80102b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:685e864104193bc3279ffcf4:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dee9853401a8cbdb728be055:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f08c439be56316b9080dd5ba:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51280ec5eb510c0a97afadb7:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab2ce9be62657279cb62c903:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:511e9000cd444d45e17ebce2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b912e73239338bc21b705ab9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb133102174bcb7528b070a9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d56d6b7e687b0ccb678de193:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ff07fed19d0c89fc5f1249d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b5bab804434183269879087:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:84e155c3fec8100bdca752c1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d54d43a30b0adf2a51d7aec:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a10059fe1b4256a9ebee490d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5d2835e70cdf0207eb971a5:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 293, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:86124d379be3e4413450521d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 293, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:efe142c7a8eaec77652d4306:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 293, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6298ca633d87a99776608f29:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 293, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3cd6558cf84d7a5994a6622a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 292, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d76549b65f6aafcc5670f615:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 292, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80febaabbac7480b645587ff:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 292, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd157bd523cb73feed988157:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 292, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c667fbd78f30442caeb4e97:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:08e62a44f1cc4060da22ea9f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa3baf8ac48d9ffb0b7ff915:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:929e09f46361fa24a970c282:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:64f46f6ac1a317cf8acc09e4:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:09756350b0ec681b3a14174b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:88ed4ec705e6a442cc3f587d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50bf33865a0deab0365f116d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:37af3284cc151a723eabe51f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0f112aa4f4a68cd7eba0c37:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:57a0ec0dad63bcf13206c0fc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65c9e08f92a9ce202c0bee44:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51ce8ed8a12ba504bfe2ab70:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c915e3987dc51091da144fc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:660306ea3c618a2c8d964ba6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51a70374061587647b530d28:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:323443dec2bfd8fc64c627c1:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea9b832e0022f5dfb2fb3c3e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bfe049b0a54b4bbe62048bf3:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07cc9ad2dbc6a87688551d5d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a47f17d3591dd007fe483f10:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23f5dfbe1d64b1cf8a522228:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:abe20fc8fce53fe9888b9c1d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5e60d2ddbb6468382c08734c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a80f96b87e8d723b92a77a10:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39919d49c9207c1e90f5a595:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:295a6fb411ad8c2a8ecf7d84:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23786cf4761e9b4b10e9cd4d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7daa4c4f395c335381c0d6d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2534b7357efb6be74fc48b7d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:798964b71191e9362595bc91:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d62d16366ccbd619783b8d59:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e727ac0c846d35db48a16604:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67cd960e6ca8f35adca973db:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a046ff2f9d09d7c2c33aaa20:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f8e25ac2ea9d85113b86073e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed206f27d5fd98e4879c9b15:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9aa1dbf07565bc86881ffc39:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d8c9dba76f0634f25015d92d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92d6713c5a1005b29650d9cc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:155f6fd14641a2072ebe821b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9cf6b23995764fd03ee7bea:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dfcf1c40c9a2804319bb0b0a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28654d3ec6fb4502546b014a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:03f2ae04e713ab4a0d75ceab:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fabfd8e25fd21bed19be5b3e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:63e1aae23481dce0a89b5b6c:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cad7109adbe1d0a97cb90b64:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:94ba63fc6f08206a978c264f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7299a5fb077300e87ab688d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d54ce421f99ef307f3f59ce:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61e2ab96ceb285917e2797d0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75aad09329a738f0363ccf49:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:de65158bf80d5a38663dcec9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f768955b02feec7cfa83e201:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2aa90e678f7c83ea667bfd4f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7835a32e046860a028c75089:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:579cf9fe496f94d8a2ee08c6:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed75ec25f215b211bfbed7f1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:536db834b5b818bb16daa79c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a36bf8099a156adbddd2ad7a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7f3d2e4a645d95404112b30c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c0d19e415b0bcb3aca49fb9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89779b3d6e87b2f15c2de4c9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d75e142ccb5692ea83e2aa4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8010f5ae742ae3fe2ea75565:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b05572e331ca11f13cb393f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bc5f6892031a4c76ee30c1b3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:53a373b84b65c14efd0a4836:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d789477e68b50e48e9a3f79:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:610f6229e6cda13d25a0b7ce:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e6a6bac193d16547aaf1849:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a41c002fcdd40680188ac946:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0241eb9082b557b013bab311:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bb89750127ac7be3b25d710a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a997f65f7e5cfe2a5a449f9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:84c07f9f155dcb320c251d6f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1691659edde0695028e9dedb:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39f0c8cce80e33915eba1af9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbc4c721c2d6cdffbd79c471:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5e5fabcf3cae0eeeaddc500a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18cc15219fd4a1b47cda8924:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa51313c6a41f885e17b07c9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6646e1db106cf350a236e6f0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a6b65f9da451095a51adaf1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:110e5571dfa4e32984d8bbed:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a13077063bb4b2e6453f258c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:de9461d82cb7e70d2efc9a9e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d6315c1652269686febb597b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65ef4deb55163817323437c5:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d07dcad3068d78d7e7caf1c4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1aad6d1c910112d41557409:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc8ca6089c697bc5a1492ca7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:545e4a8ad127d85f7be2b1a7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9324a60519705f2e6ec6f50e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0084135d34eab3a7ddb40dc:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c16e7b978ef83f776bc0da4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7384defad0b09055a7e5af93:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6e9827e301e4b6b4ff519bb:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ec7025b451dc2d96d979b36:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f57147da19294108965d0394:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38421e0749191c6415985d5a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a2d8204e9a88523541f8227:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:682351d3410d2bea2fa6f1b2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2870f64eeb1b539fb2f77edb:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a803e4987d80d76e1687195:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e56239e7fea917609bfd4e88:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f57263b7327c7058ad8556b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eaf990b375eb6070abebbc27:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:76f41f8162d567eb515bef20:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8961abd3696162e3ad8fa607:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67ace2dd0074d6e92a9d8365:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e243aa382a403df328cae226:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2530b95b77a385e6843b03d3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:084ed959138bf7612f20a416:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6387f001ebfa13c17ebd482a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ecb2114c65608c3b51e3a6d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:786534c9af01c99a755be110:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2284f5093519d1c18fc2035e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:882a6e804d22a30368d8c321:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c7750a82db748fcbbd9219f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:339a487d3651f5f4ea4d0b87:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47efb15edf8a829a6e32943a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:657b760eef76afe98564a8af:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ad7bfa9a4d549142e5e05617:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1bd284a9cdb1f1050c9a503:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c72896206e66d0f8765d8e78:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e86b247e390217db561a58a8:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8ea01478a316c176249ea340:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68b06f530c85fa183e07c6e2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4af77e879dbdf5782dcde370:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba8ce1ae3aeebe4f43d17812:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93ba3b56fec917437ceebcb8:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0dab6df61ca5ab3788ba126:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28c93c1f06a1d34b67e903a5:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81856ef4972db3d940df2c10:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5676985d18ae665f15ab884:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a36433584cc2c0e6be9185ca:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:709273537de2e62d3a52c427:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:654e7772810f847d66532c54:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f24a3617fe02c91abbe5041f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2350d22c125a175fb10ca667:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc69b1c49aad3a88ccdcb851:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e233250fb2d6462001431874:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ba78f9f3c619329d795af5b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f145d6756d87026f8b6195c6:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58792b538006a2dac978a42d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cec5bed52d30e8e43a4548b4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b74b7c08ce380e74701f600d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:510ab59890d54ef432234f49:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b67ea8d1b41d8f2d5e8d54b4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8451388bc93fda2c517d8782:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d19a66aface410f373b064c0:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c3c4963d436bb3972057ac4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f3c306cf3ae7ebe371984e1b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2529c991ceddf93e1774ca1d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddabb0077ef28acca27ca222:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a2d479f2af661bb1445796b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8cda16d57cdc7e27c9476e6b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf5b4d14e4dcf0727374be6a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b93fb286759b8e62292784a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:13353909aaacf548282639b0:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50d51c7ab2491e5f0bb3c92d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0366d43cd88b910555d9988d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee12a2ee10b48580d32943e3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7af336f36b376a6c77ca6842:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f05e1a2045f80e1d836c1581:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e71edd437a6a8a07d177b13:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f62ca6a2ed225e51db14225:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a5914416468517a0b0ff145:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d676b530ca5ed1ced22a0387:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf91d881fa5b9e0aaae9ea52:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62d6e02278abc36dc07ddab2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a017efb6234aa7b31af7d4b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:76ee37061a5405d319505fee:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:78fe39463e18193f6f722c20:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a5b7d707e00b38aea06af53f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed14cbcf4b9acf1423e88f23:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b8b1c5bef3c12a8de7c85e3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1aabc92cad14a9713c13a5eb:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf77734c437918e8f7e7731a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:74ae31ed6c8172c695484630:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bd9842ee9d27d9fa47f8266:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e3b45aa18ab1b92efe48ee5:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fab8bb97fe11f1b62e8b9b27:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a098a7001506748903e34244:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3bd74fcf796dacaca3c1424b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b9afd90d8b5580a4cf108b7e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdd5f0a1e2ca707225df9d2c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:05235228d48d2c0f6e0a02f1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3827ae0d3110802458c785c2:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c1cc7ddb7b3ec2c242a9684:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e203b12dc1b7929794a55556:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:908f2ea15cb1a58118ee639e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e3f7b7be4d54ef85f042af0:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ce606021e7d267fb211dc32:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2fdeb459ae6aa66bcb33f4fd:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:71ecc43061baea0bb8d6f17b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:605a84bc61b79647bb6befd3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a3e0719675d7b131349bda7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06dfab5f3586102c06b11bd8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e155e533cd3b40514de4cd2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:914f5793661bb16fba025e11:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c0614da34fe6bb649a6bccc:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1301ac3ce27d89bdfa895e5e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdacdc72817d55ab8b915247:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89cc9329d953fff42dc5511a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40c004c56e45ec837a9dde20:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8deb92b9ca849dc07976c082:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3cc0967c1105c10f9282b781:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1043ad4e1629659b9f22e73c:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:193473b8f7115a21adf5d14a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:86a1d90c6d74157ebf00f674:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ca5f56a805d7310b948bf64:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2fe55000dba0e9569483234:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06ffb824478d7940d6e87cee:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c74f3747bf72aa58fcc6362:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01278efe0ba68e26071eb7e3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b9c5c5aca5f7d162c33ba59:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e93cd98533b0ad7dc32081f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e986e02f7b5ac320d20e4db:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5bcabb6356565e568c2385f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e8fc97779ff6fc27ff2d45bc:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9e35df41f0ce4b91d15516e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a96974765448c82f9782697c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8766b052f48a9b2141f028ed:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df72184ed05171cb9855dbbf:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e110b6cd6d99c2b269d499b1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5129ad65df8c55a84df23027:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92b3976d280a542f9c9aef34:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ab53eb1f38044b52bae38c3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:422f9eaa4fe7c1f8a50ea2dd:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b7d11b16b688d51b539206f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef3060728bca1d4cf0f019a9:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8859b39384bb3674cf6be6bb:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c0555aaa0232ddd41d8e2cc:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3bf1133b5e36177f94ca1a0e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:707c76f4fc5aae7650e4e4fc:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:449fd5d4a4819694c23f312c:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fe60882766a2777ab2626e30:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb046d267656bd328719ec33:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:991ae9c245f3648bc6bb555e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fab016da4b8b42a7c1518243:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b065ecb382c44810956dd656:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e140c312293fc636872e53b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9157ea8f673273b0215af17:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:570df5aaa332d4a68c18bdf1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:488dcd8e14dbe73940196dc6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9f9f6be71a35fb5486446d3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:acdf1c88b4e2bef4ab7c7849:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18d213a8bae7b3bccdf613e8:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51c409d6916c6d030ad1637e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5e4f8f1c65e2a9b22d376cb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:812d78c2eb35aa32c37a597e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d0b562dbcf35f1ea7a09744:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:27afb6de88ee9ec114ec0486:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb078bdfdc6ae4277678c522:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcbb32bdadd494ebfadcfba8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d8952313ac103008ae74ad1:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e98e88936d9b8ac9043b78f8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25415ab0085df742e5084d07:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87c192cfcdb1856fad3478c3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11c13a2daff536edc2fbbe36:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f10410c1e300c2ea71a05cd0:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d2803789969cf8d09304ba37:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f5547636f20990e8dca02ee:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d95b5182df9b6f9aa3613891:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68cdf01d844d93287c153277:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:60814b3b9638b09ed683e08e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e8e8d0c6bc5136bc4ce5ede:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f2b89d0dbdda52ab6d6e50ff:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4473de6bfe4b7235222e5fb2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d56af286301a9a66175f01a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f413a7105a191f14f3fff3c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9edb6cfbdaa5a2c6de9aefd7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e59f601468f29e99099e93e8:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4f926b73ab2335f7487d643:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50af27c07d00484fb5fdc2a4:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bbb7d0605872b41c51f51a7f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b93a7bf8525b63edf9ff956f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa9ab6774d73d5d7f080f96c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c8678b7d2c720fb65344a6e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d905e9d4b72f076d33c13d7e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab6e78fe497743f0c665271e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e299c8c41ebb74c6de1b84e7:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33cc9caa2c248519b00b234d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f1fd043e4bbabd2b700e5c8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3717b8258bc68faa8c07dde4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e10ac8eb3e3daed71a532fe:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a77ad1f41898319f0632a08d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e54fcfb4598d53919e450d2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0405e987ba41594ccabf7c73:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:775e102d75596b0afdbd90db:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fe6a66db0a05664fecbf76b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9358d6026a16d685e434f6c7:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2eb3e35689abf90e8bcbf390:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:64996692610d2a5b58ae7d13:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa647d7e2e1d8ea274a812cd:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ddd764f66f874821732ae62:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc4d6270a2dbe0133403c5b6:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:628db00b8e82c227f96717e2:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb14ae09fc1f5737d251e397:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd5bca27b03de8f6c3efb042:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:512a26d67ae11e8d2c860f82:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81fd2f3cced5a25b0c96e652:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58534746f5ab71a986a774dc:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d64aac749500f9e584897d5e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e2863d514f130ed51f903779:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b72853b473ea0de9384cea26:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e0dbf6be9b6a683d91e6cf6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0cee46a52059780eb66ab99e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06790038d3806fcc21e42a9f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ff08fbabe8d5f4a1c81ec57:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:531c9d02df53393befafef5d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:169b51ac2e4533c2440c21e5:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc31237f03923622616a839b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7283b6ec02cc568c4f5eb14:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fed65e8f556072b24338fd36:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66a6d74d15970f5e754937fb:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c83eb465fb0b7fd14c9277d5:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e70d82a90ad462e10db6ba3f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d897d668231ae80605f98de:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf23e82bf53e897e161d8837:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea4c7ab25c8f096b2dae54c0:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b278426dec26bc256dda4e75:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36eb18bf9f8cf102395d747e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c0a4cd9942e23c9e23c0a05:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8653fc3c4700ab9b3b7649ed:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c8d0b8e1bc9b24f82d07acf:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1b26ee01196e1ea7f07ea3c:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:04c467da98d7f56ad3c018dc:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a6a32cc5d567d9009b1f49bf:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58eb17212bae51b923a77c58:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b8594591afe004fef731dc8:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6c5b9ef3773c7e70f81d46e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eec5eb8d0f252bca0b15e523:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83e304d67f5a1171ffc5a562:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2707bb2a79bc6e792efa79fa:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:522c14a39f5da560faf1a97b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34241971057b390340995342:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a41798edf526abeeb95b5363:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25f1f7e791ea0799b23ae83d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8420346b0e5a12ead19bca99:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:083954c9631613f6a35c706b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6af055c4464294cc4f119f3c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb7fba40afa7b36a7cad11d3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5680dcedb9099c4244e49510:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2838d4c9fc818e92adacaaca:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3803ff8f185cb6c0b864b625:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e2263572d15ab9c42991d2e4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6964a194108140010a727229:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdd5dee44a8368a4fc90409d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5732eb3e29a444c6ee65874:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7faa9327efffa3dc14583067:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:443b0c18c2875f8a8e37a1f8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:77ede63aebd73d60183893a5:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df8448608a68717291c0eb18:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5b49b6d4d2aa5d67b4cf944:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6634979ba0ea6ec8bbf26fba:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e564ae2439e35d264ee76006:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f4e7ed27454a23ea024d707:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:264cc44adff3e9f1fdf86480:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6bc0ef43c0ace87980e8503f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2307e0ee9f6e7a69fbb647ab:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fe71e38c197f8e48fef8e48:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4cd42555038fba0422650601:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98c99c306daeb5bf3af9094c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee2489584e1bfc6af99707a7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac323cf2eb1ee685dfa522c7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c33dbf8ee24c8a70b5d50045:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd558108f5d2fbcab5bfa70f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7006d38bf86be5d8a59c9742:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:523595510e7296ebb20acf7b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf39e2bf39481de8f10256c6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7f413f388cfc14001fa5a824:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1d6b769c01f031cf30a8b17:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:728be71ed26f414e2a0ca7a5:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:45ae094c037d5fe0704671e9:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:891e1b52cd7ef73071aa5dcb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd3ed5fb89e2c15c468a3fdb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1299983abee789571c9ac385:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f72357ec9461fa7521d7c9b4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b86a8e0188534972a6465778:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f685189d7a3f0217b9ed8fc3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40e67a322acc3fbbefcdafac:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8355d71eb417a7a94fd9f0c6:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25620227f87474990b837326:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba24955414be2a069685756c:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1a8ac0e48da71859fc97ff0:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9806f7229f492f708ead3da9:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c3bc051e04ce799f18b5654:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eedda693c65651a283918434:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ae764a9ca349c92d64a86586:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93f1ce93859b7cc9c5d8a81f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b975b51c04f7f4696fd7920:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c89642c2a50b9cf7f18e6002:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4148e4dc8b3a5e56083fbdc4:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbdaa8face17ddfe1a98b26f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2eb0e0b393805ef04c67907c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4fa8143ffc721f6280353a0e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa26d6ae17994438f6c4ed9c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:72a98a6bbb53a9f2fda94f60:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b67e800ba2a3e0a6eb459272:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35cfa778c07a952c2859155e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a43da79e448fa36e76ee3d1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:654ce403dd7e0f274a036ae1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a0020baa4fb513918b26339:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6254bf5d53b4be7132d2d910:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e39ea51203e2b96e7dd12341:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a55f8d812b545a871dac869a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d43103db8078f4d067ad3a1b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ce3cd2bb718f5276f6680a3a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4f342a36cd26224a43d1bb5:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:099e0c7417dad965c8f917f2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2582f7bd8825d9c4af4801d2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:03afd18817e9b25bfaa5b698:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9aa0c14bf7183cf4d1872346:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:317c4491ffba7f46d902c29d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d261c4848f9b77d95286b9a4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:edcac26c71b61ec69705eed3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ec66f54076ac77de1d51dd52:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c634002c241dc6fc71bde51a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b11664acf4121adcdc67111:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a4f737bf5b5e0901b4538c6b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:82c075aad51ae5a7cb6af5a3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35c16e6d9badab247444d9a4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:407a51e349e7ec8b16b3f1be:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f845d257f2f3270388cbde8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d48825df0e1c4ac8bef6ec00:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:912ac89969d3782530f92967:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1047e0607d8e4830663d0505:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:14a876cbff1015cf972308d6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:64018ec56babe0609839e5a9:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7253fcf0764a46c7ff28248f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0103dd7c32ef82a1beee95e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7a3acac21a9c4228a9ab0a1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:836f0a0387bf6119f675f5a4:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:902cad43390a52291cde0b94:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6a1e6a0e0a2182d6f2c21c5:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d8b2d3b72fd44649717b06ad:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e049a91a38149f16d87f919:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab090279f03bfa45f1375e79:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ece7ae3f8666ed6a9184b183:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cbf0340aecd26990e0b8412:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aadab2875c22b9b2c62403e8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 554, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a0aa98887d61798dd7a853fe:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 554, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6b30f34bddfe17b78849f021:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 554, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ecc2b752bb2631f83cb9f6cd:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 554, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12401d2736944514687c8ae7:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:48eaa4c95dd48f7beebf6fc3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cb7b1d761b288a522673f3e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4899b65e15f570ab23fac6a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f846081c095ca7beba167a7:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7bd9a399059a2078b58da8c4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:afafaa1fd9f895ced2c4c1ab:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa3b3cb3af6810aa3a31d11e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be63bb38f41a4941d553d5a3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 555, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:717d8b1c969ea380f52afe36:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 555, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f73f895a308386d08adab8ca:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 555, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4768dad76749b886aaaa609:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 555, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fbcf752aab364fbf97acf0b7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e6ea4f9d07c0de6c3691949f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e51fdeb1a3c3c840b7dac1d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8bd8787839eff57d85091339:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d918756e233ac53f641ac325:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f19b01bc274a489cc430614:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f459fd3966f63f9bd9b08692:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6cbcd7834565e3ac20042685:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f178357f12791b2cdfdba3bb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c52b7fb5554667a02023fce:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:63be0db1c2badf142dc7235d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b4651e4262601b752ea2da4b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:00e0fa31fd378d209ff756bb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab1ce7b7f1b4874030601e80:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f1398f7918a12f8c420e511:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdbe56852c5cddde7ede66b2:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6c933cb93ef12640810048f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cc5bf01719660c9e3942d1d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c66bdbc23fbc7bac9b650e74:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85dda89cf9227245bfe7946d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ab0b66376a9bc06310642de:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1329356437645d7c33293744:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5bb04b7b365ede2a6b923669:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81200c3e10f262a29d7f8abf:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38b12c8ad0af9d6083610899:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a9fa002546c887a1e6052fde:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:838844fb9c237c11b0d67901:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f74612b21e17f245b6789018:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28a4b7f5c29ba1e8701d7a6b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8bee79293ea193f85bd8f408:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c864c098b1f1dd7239a1bc3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ca1aefea5abe484166e8d84:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c79e74d67cd41d9608cd6b9:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a428124d2572b4ddc5f78b69:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6d82e1c9f5b568f5f96ba9d4:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f37d48c368380b33e429322:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9f2acfb91ca30ae535d98d8:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3db0e5eb4d0e8fad5ac9bd76:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5c5c9d6fbe0078b0b3c45b3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d62dc2b322644af0c4b1433:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6bf98150b40fc673ae6f2e06:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:050eb5903da9dc589640c7ec:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41a75f7a84558f7ef2a91791:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:786721715802ac6831afefbf:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c588bf292216b33b8b19ec91:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:573b8a5483cc63b744071e42:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:77cd96f9eb526d87db0426bf:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e60a3473dbafa2db4894f1b6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15a0a79ad1804a523d4fdcc3:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3aab322e68bc40267fc81e00:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0930f3bb683da1de730e12c6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9dec1b50aa908b6697c50c8b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:982a1f24d0672f692f3ab260:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9f263e16582be7736c59e60:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4eac891687aedc6e20a2fe4:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f5967e9bdb058ef58331e86:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9aa936effa81886e4f8d2aa:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e53af6ec6f48e3b6ef308c30:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf9e10a49a97e9e66cbedf88:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6b1658a665077870054689a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6415756613b00d63f603dc5b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a1d2ee67666e9e76a433d98:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65d2a2a53f5eec3803ba833a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1aaccc663a04aa8228bacbc2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:355688f1d93ce79dbd3551e6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40e09f3769968df2baa702fd:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b8c53ddb9c12a4fe9446e92:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c87101a4925e298aa6ebbdd6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15690dcfcbeb905525378eea:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3269f1c6715be32f5918b43d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81289dcf882179350140a4f4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd769d1328dfa811f037b9e0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5eee8c15777f36aa41d2432a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5ad4f998350d94b43e5cb66:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e6ad48460feae23637d2cda:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34c9876ee6537c7f90292664:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16347b895187d24626199382:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3f4f12649bfe560792742ba7:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a546f5f90b7b7ac01e1a4595:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab7f7d5f4f3c302115c82baa:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8d8c7d1aab200648dee1ea3d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:310c7ba7ac4719062817d591:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:afd236258599e4a58de17166:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6d80797c7424475cbb46519c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d524de9e4d11c6f16c1a4037:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc3844f44d7e3bcc429c3bc6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e99aea1e810e1e7b4c305923:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be6c3ff25d7cf8382b90f9dd:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:acd2a3c99c6af78c6e465be6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d7b6bb08c570fe4af3024659:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e11d7dcd3f35cc8663b8b81:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0bdadaea6ebd39a130e72e47:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:26a1681e7f60bc341323040a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:82203c0c43bf32f0d93c943f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c2e44d61fea9c77aa16d3be:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a9cee5dbf3a0ccad88aadfb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5d095b12fd7472d3bc8bb69:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31b9f493b908961e6811ab9d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:edd0e2ee17dd2da478eeecb4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b455ae9104c57566d6564252:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:640b059af09b84e9b7978903:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3247f454153089665ec9098b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e95e6a94af634c7109c3f246:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62b69142175d94f3fd4e2d72:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f04a0e1c43053b82d55ecd0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb9458a3614f91257f274be5:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93fd401e3dc8e507c8a1464c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a4247841838fbbc31bfba19:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a98804b3a9a9bb3d16bbe974:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:279040d0da84d719e1008a7e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5013d3f84a766ff90e1d66d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:95a8f41f1a8abd2d3a9eb1a5:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28f66441157133b175210bdc:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2480546c3f0750ab8889063e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a6bdab83a4b7ecef3ef647c6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c72598e032df002a371c721:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed030c57b48d07a26fe3fa94:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f49d1a1b8732118af3e1be90:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41337ab27205460f2133b691:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:569e1b7cb9d952a5879b5568:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e40525b20d565be581c4ba0c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5359bdc648fa50d4cac1891:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c29a5edeaf7b44106b0b134:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c367aeeba3ed47df08cd733:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:866a5c381de4ac62118dff85:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23e4fdffb6c15673bcdfa30c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf34ab76605a322f3022c5b2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9dd8042ecdbb399e6c0049f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:465eb1946dce6dec3a2a5ac0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01d94f0b21a5dd2c818b3cd0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38f2113fc226de905ea3e002:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f67e5938b2605fee54199d36:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f9fac3a5c1238a6bab5e8372:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6978b8de7ca7784877ca9c1:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b4ee082c0adedfe3e83ab0a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aaec28883f5106f66e678b38:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4b666673a2b69725f3dfce6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1dea05de416e4b30a3d9b3bd:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e3b4aef7ed11d9fe77e2363:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:539335d4d742aa862a840934:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbb73448f12660f057c255be:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69f213a32868f74881039073:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f61de138da767b3a06938b2:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e34402918bf0bcc3917c54ba:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8fc95d8e56831fb80726a1d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0809c5a5bf2a4a227982b51d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ff22df9dcf793090dd745cc6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28b2c8e61c631a7343e23fe0:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a000537d614295fdf04d0998:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:131b6f3346e6a783bda21564:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdf794b73f61d7e3a877b6da:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0a5494be9df8886d8899ab5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:255ccea5a91c2b6559fce709:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8078bf5f14220044ed9caf41:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5798b04ac0704290fc794ed:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:49dd3e24e9d0ecd0bd6ca487:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:db25e3fbe8afafe5ede4a0f6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c28d78407447277ce418d6e2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b52a2679fd113557ab42323:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb116adfca97f9e016d66bec:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7047e796c12bd31f40ca0d40:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f03be949c020e94742f9e77:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d25222ed9c8a3bd46b12cc20:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:463bd53f0721e822217ba46f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0ae468377d98732c59ce2f3:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:46479f8f0607160ac6dea63e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb2996746d4a3a8542f6ff1d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:077f815d1122414cb8415a85:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d573dd064a37b7a0c385b8d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3412ff2ac67063d787d1c17:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:462618b61651570b2846c938:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f1183ebd02f47a05a9bc908b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e73459046bdc8819f7836a29:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ae6647d9b35a7839e4ef83e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67f1d172f24a2b42d4d9abb4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fff63af9fd2bd2bca03d9d72:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:637de0b4ce60d431b3fbe6b9:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4748d544ff79abfe5fab9847:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5a5948642f75f93c2f3cb77:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac5c668d5d11be78678d4a84:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:126562d789deb80356468d41:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:57a2008078f742ab6c8626e5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a4553b77d5b53b5790e3d20:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11a4cd6425ef50288dd89dbb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c32ed72850524b6115259475:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:43b0d5e950a0b165bb220323:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f8fe2a27635c879fb8628e97:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ec1668afb0674d887d30b5cc:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbd521c43201e26268e3aa7a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93ce0e674e3a155e92ce6ff0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea556a3766a137afff9d5143:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1c350beb0696f45d0b636d0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a561199f0e30a86517d4e77:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7c26cd5a928391332a449a6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e112c4da03e89806f6b51f2:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d21883d41d3b3b5c5968af26:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b9715bdd36b25a38ad6486a8:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:820626b763f8825dd3068332:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93d08bb7e0b93cb303105554:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4a68ccc2d5e311494876e75:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d24a7c504ed2622641cdbead:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 283, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7cfdccb7443a93a634c433a1:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 283, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:817b57139b63e108f8790fdb:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 283, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d941b724bf9df118dd6a0aae:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 283, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ee3c0f4347f3f58e953611e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3ffa2df372e8f6dfb0d8c38f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:471380d453ba14d8f9c03077:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5387a48bf735045df5884247:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2082df371acd1246e6974878:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e51957d703aa7a21a9126f7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0b27fc187d5b92f2f0baf02:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:388029ea4c6b9db0b05498d7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:596dc4bbff62dd637bf1abe5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f95e6852ac5ed8170400816:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d53ad7b791e6754ee684dd46:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6c5abb646277c869e8f594f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a0e820d5880fd053a23b8742:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e24bd548aa26f5edb96ec44:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4020df98668c9b44c45f1f60:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a8f4fcc01b49eaa3bb6b30d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c8cdf2cb4c1de722f45f3936:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea686bc3ab05831b29e7008c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c938b97813786d8c82237b3:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d649dddc9b41e2ca653328b1:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07b6de7eb9bd7361c486b345:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eef957bbb536e28fe4381bdb:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:19a7d6b2323b098d9e22032b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f2aafe130fc41373b6640f18:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7611787ac567633d94de55ed:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:648bbf706e496cd9f6908242:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0cd71dfc18adc305ffa5780d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:262929de5f678d99fd81523d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1cd8bbf14e4a59fe63113214:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4402a299cdebbaf033c9050f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f882ed980b69673ee6a7b58:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5d8322afa755ad70d59d60f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b959e9bf77a5adc5a30d2ce5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d189f2df0711b7fc2c771265:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c16014d87823dd357d1456d8:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:128f395ac37d7006e4e4c367:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c29119aae7c02211285f7082:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:697e27e971a1806c494ac95c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8047d9de8466175cbba13672:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5025ba4651167be2693e5f6a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f63abc827e81e246ac37e965:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a80ed0d47a0d3026ed1dfbc:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e7fe1b673f6177b0fa1dc76:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b1a6bb2f1ecd4b535f679a07:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a50961b7fd5d31f3923d1b7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:441340c226bb1935b121e41c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67f031bfe4e9e92d197132e1:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e0babcd3043ac68136a197e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5e495780c777233199b2eee:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8bd553e22df1683db52d5e78:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4892ab35e607d1b27a490c56:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a5f71abe56130d9b6d00a9e8:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67be1599e86d96b0a79d4b3d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c8656d105aa907f177cd2e4a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:501c446c78491b8f0a6b7edc:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:db5889f39922205969c462aa:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a67f9b5b6935e1e6b2573e9c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5b299de7d39edda771cb7973:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:342af4ee63b0cb125b271a4c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:625175dbc051d9cac87f6110:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ac053036472f06ebe164a62:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51f4549cfb5680161d097026:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6231a699832f887534551fae:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ec98f845e66e9a7221f2312:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f53aa74c3869112597b5fd30:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b13b8e3553d6936297b72580:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92cd14a08b1204b3c59807b6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d250f37f1770efbf1655af6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7bb180e591734d11d40b8605:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e7adbe36391fe00ae9a4164:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a73aabb2b12d1a995f9bcc66:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5978de20565039401c0f9ac:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:814557a8578a6d65feb6e938:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35e4e09a850e616d48db8bfd:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33b52c598887d9d5905859f1:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:488ceed82540659725d8c36e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23c3f398425fa7ae16f59f98:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b7eb7993227f08865ceba7a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:851cc9d5fbda6c0c718c4fcb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:111c2d37ffdf960b57346949:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c79b07c72f1575107a85cd9:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d41feeb008c6f9c5b2bacc21:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:896b4776fa05a557779d30c4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6ac74dc2de34b129cb9155a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:86965a2f8f90ff782961f060:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cef8be7cfcedc847fc2cfbf5:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5469358631eb0c0fba97e23:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31afd4f6e7c73256fd123f75:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50e9b66e55411fec32bf17d7:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e8b47dd39a6ad545fd84c1a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e915c93c80ef2f02d4d2c4b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8ff178390db9980396b19358:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b6008f12b6d461b6f5ad121:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:55057bfba949c6ff0663fd0a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fcf6ad681a5145718215a021:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:428aee950ad2683c813213a0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b91f11fcd38154517858dc2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8259f446982e956a0ca11695:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d36a15fe27e140b6370c6f7d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:637cc81e55d6b29b3859c430:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8baf32af0f2efafb0884ef1:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be4b7e8f2f86f141c2d8d3d4:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51275459581e4b1ef3b00b7e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa2a4a1442ae92254da80c38:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbc1dd0f7c711abf7adab9f5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f8bb96e5dbcca559f3dec2b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4abdb09fd01605d136e6fb41:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:806d0448919413ca550c11d0:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8be2af3bc78bacaa2c393ca1:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ed67e4159be3dfc4fbe7482:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35fd0ba5d43edd45fd72ccd9:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9bd3410efc8104bc569c0c73:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f7fc6e2d78bdcaa8ebb70d1f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 588, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3ae94e3a3615b606c6b0faca:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 588, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:253b1f5b5850bacf600b161c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 588, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5ce3b5ce3aa46d05142aff02:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 588, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ce51cdfd50612eeba18aa9b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 587, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0138be8aae7f889727aa1d1e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 587, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b7e759881886577dad273fd:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 587, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef6acc71d3fbb16066661281:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 587, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:27d8757c9da7f3430e5e0ced:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:194742a1a19aa36e85a58c7b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31f3f29c21bbee6875a62524:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d7621d7b8893eedbe393db0e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4a3eb5c21fc375b3e82b339c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d10cb67b66d9c0fd81ac7e58:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1848c23cd4d35898d022bdc2:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a30ca432414e10d4fa53896f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6c30c88fd3483ab14dbaa1e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16b40e61859c5318b4586afa:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a0c1d6a2009f8c9b6913994:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a0bab24a3ae166837ddf068:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40ab5b0e677bf9770e81090a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:78f00f7c7878440a2410f788:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f15638d409cc44fb52c8b7e4:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1131e5a097ac7e10af06232:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c2b532fd0a1c316f6a7827b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:def4366ded6a6df896671e94:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c25d1c92537c287c8d1f9ae:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5bb8c0d305fcc9190b60bd0d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3529da347672ddc6efcd7cbf:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3eeb11ec8c5b578dd385f823:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:55f8f4450093cca5395337b4:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4845c32b3bc94002c70ccfb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c29d1885fb3e711832008e6c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:501e2691a84e6ddeeae2f9eb:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f293099110937a1af731b7c4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92488c2f6185ebce0ad9f995:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a07bbdfe3071266e2b3b2ca:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6595a8344e600006d2123f9b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c4a4e147f7b1588ae0bad92:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4979e947e876782bb7b71363:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd3f7c9444d9e11b215839ce:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:866d5c488eed87d63b94c01b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b60f0028ea5cc952f66ae471:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7279558be989c68b7c5da9e7:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ec625bc4ecb9c56c9fa0cc5:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e4a7f385594c24fb238ae4b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6f2292977793d8d3b0bc2a0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e331f261aeb0da55b58b86a9:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e4af581b0e4f9fad15de3b2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:102959b674a96c013936409a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01f5cf08902258a28cdb1212:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6cf9cc6e93265277a8e3ed0d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f221f191c4d26720283f4d1f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e11a42fce820771511feaac:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15450d6b94feffbbd82fe3f3:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9e0925630cc8108b1dbf766:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6b98f2b0a9e5e50d70c9e675:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:241aa15d6210643b37819f49:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c8b4b84f80a8b426b676149:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dfec15d15538d2133f64505b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:332c305d9f9064cca7c91b1f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c9d3d9f73d49180fd4e70c5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8353849f306c99b801e76cbb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e5c16ccc37f603aaee2e0e8:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5af1d3651ed77b0e4f433bb:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7b1cfff32e388417d93f789f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f3fc69214f96664fd12c2681:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f10ee0a7b7901b7b0cffb768:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b1460dbef236498d111bc421:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1776662fd2b3d4233b7b1f84:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bbd1fa1192f750484e786f94:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8ecd72c6f944eace26d5a99e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a5dbb5ddf4c974b338731c6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2afcaed2770a429bdfae7681:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f0fd919b00922f0fb8004de:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91493ce1dfd1c9efd5084510:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c214eb63a496e85f4e8f0331:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e9860c60f475cb0c1bbc1ca:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b31bc0c29ed3b7f06b10dbda:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c73bfb4d052e52de134e6f20:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c31a0ea252ae23f905fd135b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:305a069109d09eb613b4548f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e45ce175a6959ad79175b27e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:656489c3fe1025652a510de3:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91f5ffb49f77fc333df264ed:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66e0e51ffb6f3943c3015dd7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c12252b469789af7f5190a79:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:563ecaa55e9433ce2d40fb0d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6ef55093396c3904f96ed65:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:924e4b96ca5999023e35d70a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b8e9ced855b770c0618a521:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bfd687a18b43c9f071ab97c9:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0902d059933889080b0053c7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee9468b38861c78cbeae36de:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c77559f45064cd0b43746d8:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85244a117edda6d6143e169f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65530be55a82ba6c9e4e0535:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e11bf957181a9d4a1bdca5c:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6bb216eb099bbe2cb415eb1b:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c2a1c77b0784c41b332a688:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa2b06bff614037acff1d512:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e9a16be72ced2fe22e4fcd4:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7df43449888199a23a8d880e:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d7067b9c047e1fc87bb947d:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8f52a5c304a45097f71bf97c:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:63d57a7aeb929060cf51188e:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0ed1675348397a71ec61d8b:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e4f1cfcee5f63659228e6f6:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b53b47523823292cca2a20f0:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 801, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3bf115a8cf096d0947013ea4:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 801, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b0199d15491378b758402f5b:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 801, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:572f0ba1d3ddaa3fa4ac6f91:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 801, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1e0301f1ecf0f1847123e18:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:040fc74e25d32f5a1706d9a4:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12fef32144a2172110026e00:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:249c8645ae2858840c0d1c78:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39b9d4042a63f6ee6f7c2466:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac51c42652a70e790e7111e4:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23dd18b5e052cc51e1ba4052:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47f13308a31f26e9f85206a8:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39cfd2cf1e0faef8f4c0c05f:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:10589703550d91e77b38af7c:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5ff0bfd990d8f473981610b5:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e3d2edc1d02debf8b325bca7:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:90ead2aa54f0e923806f9b7a:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6266d716e135c92d2241af78:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:04100e1e91887aecab329285:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:640310b1795eba44df0405f7:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:823ed2dfef942950a0fe052d:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89e8e93ebc09ab77a2d12ee1:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2f03a7e02d4c7a0b37a5c12:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8df3da6709753e818e475063:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ec0ca15717c6beba10df0969:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34e19699b6fbeb555cc4c3ec:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2498f981bea7f1ea2f94f562:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1181ce3fb218aa0ae26cd2ba:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80ada521c29562b2d79f27b5:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4d3051156c025b3e9c13127:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b17e8036bb65af3443e673ff:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:858226223a62dcae09bbb797:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8be9a77dcb5da3744baf8878:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75d8ccffbb0e46bdf88c8778:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5eee2c4c2557caa27de5ade:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25092695911e018291bee049:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:249580a78256b5fd65542bbb:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:297c6e026ab9cf238b3c1d10:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b91dd50087e2e74714d9cf1a:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e56f2ee1a2eadd1013f0e3e8:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:942cd8d852efe24d6c5a4dd5:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:334d03453308ec3c45e12e23:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a267b801453dda4069a69b25:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98025ad456cd47b9148284a1:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:90f596b7d7aa2159085c2b63:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:244f73488ae8b18c97afc783:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15eb4c3530366489aea10f56:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f0ea261f90dcddea760524a:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:db92f0e465c6399c5c4abec6:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:af3cc83d4f3621aacab12778:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e2293e4fd434a600da4d0bc2:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd59a2e1e889f95b39067f66:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4997c1961582643cdb3747f:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:197e69acca81ee3796ea18df:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b58c019eddd3d2c33f22959:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ef468fb72ac4cb64a81acb9:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:78eabe08f32e049a8cb4f8b4:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:48823570051dca86c519667a:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:27243d43f64937e1eb9b625a:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a9c17041a8da3f8c51d28132:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f47dcdcc829ef63b4c6c5d0:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f4459f112d4feb1760b4176:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:de191bbf708498615876e957:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d76351a595e7c5c928de2ae:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06c1c1ac6665eca225c2382b:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fe46a01198229bfa376dd10e:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e064899b758fedf86fa2005:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4318d2e14a9b1ce3c16c2894:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f066a2f4cbc8579afd9f1982:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c33d8efefc0fc8118834420:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3bb9967a166f17cbcad10f3d:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9627aaf17a3806f0aedc0837:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a252aaae386b04fc387afc9:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38edf84bee3e6ad9c57f203e:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c962093142dea21d0ec3dfd:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ca03f0754d5ce00a62bf0ca:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:196f3a6bd50e835d24bda265:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2eef5b738aa0b07684bc1dd0:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06f6dcb92ca8f002d5047baf:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a477e556b4abb5cd224ffc4a:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b4683cbe2ce2d042e8341c67:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:05da9714c27d666caaa2a782:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58ef310a92cb4b8b9b6a34ac:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:539c270926beb6b3742b164d:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fab0fb117c6ad4e0061ee617:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6df3c133fec6415c08fe51dc:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:753c524e8accedbf5235e6aa:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a255272ba51fa1edc1aa73e:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ddd679c6df6a0939e230453:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b23c7e2e9a40cbe74d0d0aa:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca673d0199005992cdffce6c:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:031d1db0d38c30f30506a4bd:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:45b3799fabd153c39f7b6531:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8f743a364012858f43563b93:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9229baa933c2fe26956e93c7:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be37856b3136494b863c28fc:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5b1cfec010f249f02c487e0f:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02a0a62903ef89c49f64ee06:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7dd94f0d108bbb8627c3b1f:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8467efa8c94001e51b138cba:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:045010ec88340767d1b89b2b:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:13c6045871929ef1628066ba:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36d14866217fafab8eae7780:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e7ed7df6d8739bb28b8e150:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b49b03914fdf279abf49926f:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:141a5bed53a942c0406932ec:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea6882511a6c6fdde770fc68:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c50529b6eec075502f964315:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:44bc47aee0dd3387580fbbe6:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ebe5cf8672d9330665e9e010:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:538db869e56e44d7178b3be8:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c841d71a8f7ea07ed3b92208:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5db15071608cdf33b0e85961:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d69cd166660da942fcd5aed6:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:97127e2f57c841d95a72d242:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33f2a2f998988627ddfa3f98:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + } +] diff --git a/variants/events_ce_seed17/tokenizer/chat_template.jinja b/variants/events_ce_seed17/tokenizer/chat_template.jinja new file mode 100644 index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4 --- /dev/null +++ b/variants/events_ce_seed17/tokenizer/chat_template.jinja @@ -0,0 +1,89 @@ +{%- if tools %} + {{- '<|im_start|>system\n' }} + {%- if messages[0].role == 'system' %} + {{- messages[0].content + '\n\n' }} + {%- endif %} + {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within XML tags:\n" }} + {%- for tool in tools %} + {{- "\n" }} + {{- tool | tojson }} + {%- endfor %} + {{- "\n\n\nFor each function call, return a json object with function name and arguments within XML tags:\n\n{\"name\": , \"arguments\": }\n<|im_end|>\n" }} +{%- else %} + {%- if messages[0].role == 'system' %} + {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }} + {%- endif %} +{%- endif %} +{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %} +{%- for message in messages[::-1] %} + {%- set index = (messages|length - 1) - loop.index0 %} + {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('') and message.content.endswith('')) %} + {%- set ns.multi_step_tool = false %} + {%- set ns.last_query_index = index %} + {%- endif %} +{%- endfor %} +{%- for message in messages %} + {%- if message.content is string %} + {%- set content = message.content %} + {%- else %} + {%- set content = '' %} + {%- endif %} + {%- if (message.role == "user") or (message.role == "system" and not loop.first) %} + {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }} + {%- elif message.role == "assistant" %} + {%- set reasoning_content = '' %} + {%- if message.reasoning_content is string %} + {%- set reasoning_content = message.reasoning_content %} + {%- else %} + {%- if '' in content %} + {%- set reasoning_content = content.split('')[0].rstrip('\n').split('')[-1].lstrip('\n') %} + {%- set content = content.split('')[-1].lstrip('\n') %} + {%- endif %} + {%- endif %} + {%- if loop.index0 > ns.last_query_index %} + {%- if loop.last or (not loop.last and reasoning_content) %} + {{- '<|im_start|>' + message.role + '\n\n' + reasoning_content.strip('\n') + '\n\n\n' + content.lstrip('\n') }} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- if message.tool_calls %} + {%- for tool_call in message.tool_calls %} + {%- if (loop.first and content) or (not loop.first) %} + {{- '\n' }} + {%- endif %} + {%- if tool_call.function %} + {%- set tool_call = tool_call.function %} + {%- endif %} + {{- '\n{"name": "' }} + {{- tool_call.name }} + {{- '", "arguments": ' }} + {%- if tool_call.arguments is string %} + {{- tool_call.arguments }} + {%- else %} + {{- tool_call.arguments | tojson }} + {%- endif %} + {{- '}\n' }} + {%- endfor %} + {%- endif %} + {{- '<|im_end|>\n' }} + {%- elif message.role == "tool" %} + {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %} + {{- '<|im_start|>user' }} + {%- endif %} + {{- '\n\n' }} + {{- content }} + {{- '\n' }} + {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %} + {{- '<|im_end|>\n' }} + {%- endif %} + {%- endif %} +{%- endfor %} +{%- if add_generation_prompt %} + {{- '<|im_start|>assistant\n' }} + {%- if enable_thinking is defined and enable_thinking is false %} + {{- '\n\n\n\n' }} + {%- endif %} +{%- endif %} \ No newline at end of file diff --git a/variants/events_ce_seed17/tokenizer/tokenizer.json b/variants/events_ce_seed17/tokenizer/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2 --- /dev/null +++ b/variants/events_ce_seed17/tokenizer/tokenizer.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506 +size 11422650 diff --git a/variants/events_ce_seed17/tokenizer/tokenizer_config.json b/variants/events_ce_seed17/tokenizer/tokenizer_config.json new file mode 100644 index 0000000000000000000000000000000000000000..9a87e3822403fb40a9524bb75edb66c6f0585750 --- /dev/null +++ b/variants/events_ce_seed17/tokenizer/tokenizer_config.json @@ -0,0 +1,30 @@ +{ + "add_prefix_space": false, + "backend": "tokenizers", + "bos_token": null, + "clean_up_tokenization_spaces": false, + "eos_token": "<|im_end|>", + "errors": "replace", + "extra_special_tokens": [ + "<|im_start|>", + "<|im_end|>", + "<|object_ref_start|>", + "<|object_ref_end|>", + "<|box_start|>", + "<|box_end|>", + "<|quad_start|>", + "<|quad_end|>", + "<|vision_start|>", + "<|vision_end|>", + "<|vision_pad|>", + "<|image_pad|>", + "<|video_pad|>" + ], + "is_local": true, + "local_files_only": true, + "model_max_length": 131072, + "pad_token": "<|endoftext|>", + "split_special_tokens": false, + "tokenizer_class": "Qwen2Tokenizer", + "unk_token": null +} diff --git a/variants/events_ce_seed17/train_log.json b/variants/events_ce_seed17/train_log.json new file mode 100644 index 0000000000000000000000000000000000000000..d9e00e8db3945bf1fd9a31cfb29b075e3b806055 --- /dev/null +++ b/variants/events_ce_seed17/train_log.json @@ -0,0 +1,964 @@ +[ + { + "step": 1, + "phase": "full", + "loss": 0.6698931157588959, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 1.206429815851152, + "batch_question_ids_sha256": "9815e3084b0060ca1bf3f53e09caa2e8ba2ecfbfa18cd5a90c1b6ca1dc29fbcb" + }, + { + "step": 2, + "phase": "full", + "loss": 0.6899416968226433, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 1.9398853788152337, + "batch_question_ids_sha256": "0766e00bccd0b9d63535226a2a7454d88d591d6a1adb32befcd64d40174bd87c" + }, + { + "step": 3, + "phase": "full", + "loss": 0.4910094514489174, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 2.678210761398077, + "batch_question_ids_sha256": "3eb8fb253ed8d59b98fd13fb12ae3aa08473ff98bf07f26c7537d8d4c6999318" + }, + { + "step": 4, + "phase": "full", + "loss": 0.47546009719371796, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 3.3923207791522145, + "batch_question_ids_sha256": "b72b9555459b9660b8cfea9c061b0d69e21b213181cf19973fa9312c1d740e4c" + }, + { + "step": 5, + "phase": "full", + "loss": 0.472549706697464, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 4.11856244970113, + "batch_question_ids_sha256": "851370879eecc2bb2b872bf9f26a90b0c422dbb910faa7ab99625462dd70126c" + }, + { + "step": 6, + "phase": "full", + "loss": 0.47603993117809296, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 4.842001353390515, + "batch_question_ids_sha256": "a44ac9fef40754316a48915525e55ee3c8c5f613d2a3a206a70c53f6a360d6a1" + }, + { + "step": 7, + "phase": "full", + "loss": 0.4351997897028923, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 5.562285288237035, + "batch_question_ids_sha256": "dd15d5df027d50d8bc83cc139ce58f36adf2613954b88024320afffaf4f275ce" + }, + { + "step": 8, + "phase": "full", + "loss": 0.5573603184893727, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 6.283805459737778, + "batch_question_ids_sha256": "b16e21ce00030d172555d2ed47f3935798c5293130655f2d4bb114fab8d4f82e" + }, + { + "step": 9, + "phase": "full", + "loss": 0.25312627106904984, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 7.027102201245725, + "batch_question_ids_sha256": "db13a8f636762d4e5c56f60dcecdc4a04f7acccee36ed4846fe242c110813e33" + }, + { + "step": 10, + "phase": "full", + "loss": 0.5761465504765511, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 7.779015112668276, + "batch_question_ids_sha256": "d29a2ab7bae9372a6460e6d41d2320b0a13b55633fcc21e2a06f06a64a2b8f22" + }, + { + "step": 11, + "phase": "full", + "loss": 0.4777762070298195, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 8.493401571176946, + "batch_question_ids_sha256": "e435c49e7cce06b5c387742c1cf2eae1fb3f57349a03ff3229713bf8e9179b7f" + }, + { + "step": 12, + "phase": "full", + "loss": 0.7341089807450771, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 9.231942504644394, + "batch_question_ids_sha256": "52067a44724ec88a3f85cace2ddba01f511bb396cbf27877697e85efd4f0c68e" + }, + { + "step": 13, + "phase": "full", + "loss": 0.3001631014049053, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 9.975338879972696, + "batch_question_ids_sha256": "d01e765c3bf9b2b1edf10d82d3b784f51514a21682d040badd6d9c1c29c3069f" + }, + { + "step": 14, + "phase": "full", + "loss": 0.43069224059581757, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 10.703695585951209, + "batch_question_ids_sha256": "f7d76860a7e7abbd079278179b2a6054aa8b1bf90c6e38db91f414cf906eff6a" + }, + { + "step": 15, + "phase": "full", + "loss": 0.5127209424972534, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 11.451611735858023, + "batch_question_ids_sha256": "1d089d97e3e6836ec1771dba1776d38f6a8ca414f96cc3c25fd0053ed3e27a44" + }, + { + "step": 16, + "phase": "full", + "loss": 0.3563181236386299, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 12.173952360637486, + "batch_question_ids_sha256": "abd204307499e62673daa334cd3a4533448a0ef971b517e7acbf23783905ae91" + }, + { + "step": 17, + "phase": "full", + "loss": 0.36245644837617874, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 12.8896827660501, + "batch_question_ids_sha256": "6d4029fa8044653ff0ccbe80e4371cccd55b6fc7c6462fb81d77f565bde1e2f5" + }, + { + "step": 18, + "phase": "full", + "loss": 0.5653930976986885, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 13.62013566866517, + "batch_question_ids_sha256": "88ddecb0e43b996acf8e321896454ceb6e79095126d68380cfdc7c74ba66d04a" + }, + { + "step": 19, + "phase": "full", + "loss": 0.5002457201480865, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 14.346520440652966, + "batch_question_ids_sha256": "3ba3b72125f7617c5cb4dfc031843bc10773d8aa7747ef6eef76ec813abc7a11" + }, + { + "step": 20, + "phase": "full", + "loss": 0.6882804483175278, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 15.069504925049841, + "batch_question_ids_sha256": "a8fbfc400a3dd61e2468ddde908d8e6914e8742e66df75fa209575bbd49f674f" + }, + { + "step": 21, + "phase": "full", + "loss": 0.3362383544445038, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 15.863264437764883, + "batch_question_ids_sha256": "49a7b1f6bb6891cd06f7c495942bbdbd98cfbd7435f49d77922ab364b9c1c5f5" + }, + { + "step": 22, + "phase": "full", + "loss": 0.4032861888408661, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 16.61500151734799, + "batch_question_ids_sha256": "d93e27a20a5666b8fe2ea5489dac0a58f99a770bfb0d4ee6870d36621ed78dee" + }, + { + "step": 23, + "phase": "full", + "loss": 0.5400929227471352, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 17.49686106480658, + "batch_question_ids_sha256": "88845418c486e4c7d856761b8aca680a7ca52ec59c84219e708745af76a445c2" + }, + { + "step": 24, + "phase": "full", + "loss": 0.4465572461485863, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 18.21982985921204, + "batch_question_ids_sha256": "b7fcb8d686d0a87ea4fd6ded60d595c7ccc4b132fc739e047564d72f5944577f" + }, + { + "step": 25, + "phase": "full", + "loss": 0.38778524473309517, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 18.957347061485052, + "batch_question_ids_sha256": "b2e6af990ea722e6cb276f57a736c23766306ae29b211bf00447d868b434c962" + }, + { + "step": 26, + "phase": "full", + "loss": 0.46882137656211853, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 19.701538900844753, + "batch_question_ids_sha256": "67253956a0835f33db1fa53443fed54476a027338e0ffdd7726dcb48fbef1c9a" + }, + { + "step": 27, + "phase": "full", + "loss": 0.4425557181239128, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 20.427865900099277, + "batch_question_ids_sha256": "b37cd333ba07e45e13875704fb0cc42b2f44fc5317737590bb27d88a7ce61452" + }, + { + "step": 28, + "phase": "full", + "loss": 0.537452906370163, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 21.164012663066387, + "batch_question_ids_sha256": "d2a5a95e218ea89d7469dfdd8ff320de68f88ff52171675c0ccf7c9c3d95a891" + }, + { + "step": 29, + "phase": "full", + "loss": 0.33650635927915573, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 21.895237404853106, + "batch_question_ids_sha256": "1c6aa40b463139c53b7a93df96c3a1db9c53fac3123be7e8dda4fc8f53bc7996" + }, + { + "step": 30, + "phase": "full", + "loss": 0.18629102176055312, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 22.596429711207747, + "batch_question_ids_sha256": "be09ca5b88c61e0617249157b2d75545b12e403abaf2ff570be1e6d30856d742" + }, + { + "step": 31, + "phase": "full", + "loss": 0.575798325240612, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 23.33407491259277, + "batch_question_ids_sha256": "08f4544d1dfa624e78df3609c0059204a16146f8421bf85c14d0b10b5926b448" + }, + { + "step": 32, + "phase": "full", + "loss": 0.5029938481748104, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 24.04410624690354, + "batch_question_ids_sha256": "f8a0c95d4ce5a21d3b7bdd381eb7c9d15a251da10b33f0a8521a3e81eeb1450e" + }, + { + "step": 33, + "phase": "full", + "loss": 0.3700496293604374, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 24.790576100349426, + "batch_question_ids_sha256": "4c396fcb15e0a31d52d00473dc2f3bee3767da23146fb5038b6e63d9ad5016c8" + }, + { + "step": 34, + "phase": "full", + "loss": 0.3722036397084594, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 25.522674680687487, + "batch_question_ids_sha256": "c59ffed718f7ded62b1028779c65e8759de67070b26019e8642f75dd2e43a512" + }, + { + "step": 35, + "phase": "full", + "loss": 0.6066184220835567, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 26.259713889099658, + "batch_question_ids_sha256": "0e54ce4db6221fac98c15ce77d50701153708881dda86adbcb2776637215c9bd" + }, + { + "step": 36, + "phase": "full", + "loss": 0.7472243523225188, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 26.99773812200874, + "batch_question_ids_sha256": "24c257b893015e139aaec823c9fa8ed630a7d2df239552f229ebb832a7746723" + }, + { + "step": 37, + "phase": "full", + "loss": 0.5684906765818596, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 27.740370864979923, + "batch_question_ids_sha256": "bc22d73d0ac85948fd9bbd0813992b093f9beb230652c6234447dbbbfb4add35" + }, + { + "step": 38, + "phase": "full", + "loss": 0.36942679062485695, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 28.50852065719664, + "batch_question_ids_sha256": "6e47f7951fb42ce413ed2f4ce54d8832b534a3e9d0f2e35829a1608286227ef4" + }, + { + "step": 39, + "phase": "full", + "loss": 0.3635300546884537, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 29.22664304729551, + "batch_question_ids_sha256": "b5bfa6d3b4f6544cb42bc8b82b7c88ad3430367ce2f47dd0c3182d7f55e26796" + }, + { + "step": 40, + "phase": "full", + "loss": 0.40619706362485886, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 29.978176869452, + "batch_question_ids_sha256": "e356d8692763396b1f305f967bc42bbea222b3adcfd7355c8ffd489d7c4a8702" + }, + { + "step": 41, + "phase": "full", + "loss": 0.5313360318541527, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 30.714267833158374, + "batch_question_ids_sha256": "eddf9bafc965ad35c8a9189e6a7b19eb336e5f2736ab490e5d0f7547f885b883" + }, + { + "step": 42, + "phase": "full", + "loss": 0.2942931465804577, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 31.4632102586329, + "batch_question_ids_sha256": "93f507828c7fa8715902d87edef912b05fe2920c25f400745b5d4b1835ff1fcd" + }, + { + "step": 43, + "phase": "full", + "loss": 0.43418202362954617, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 32.20179224293679, + "batch_question_ids_sha256": "50f72e83de9b1b9daf154f013aa536b1454d8d8c13b503ee6efa9007c721faa6" + }, + { + "step": 44, + "phase": "full", + "loss": 0.5574786812067032, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 32.9865256017074, + "batch_question_ids_sha256": "71067cb1ca0d666b5a16e53227ae1840e468c0c39e879ef8004ac73520b396f9" + }, + { + "step": 45, + "phase": "full", + "loss": 0.31117148976773024, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 33.79001239966601, + "batch_question_ids_sha256": "b9270e713e4fca02ecc7205c7ac1e320d3c87286eb7237ed0b02a4686b626bc4" + }, + { + "step": 46, + "phase": "full", + "loss": 0.4900943711400032, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 34.549148452468216, + "batch_question_ids_sha256": "fbecc4d2e806b5b42f4f82a7e99465e1fceb5b7b61dfc3cd167654deeef897d0" + }, + { + "step": 47, + "phase": "full", + "loss": 0.35709232836961746, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 35.28404710628092, + "batch_question_ids_sha256": "9fb1e3f8e3eb222e0e377eeedad5b0df90d11389465803b67a9b5090d2e79ca0" + }, + { + "step": 48, + "phase": "full", + "loss": 0.3858681209385395, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 36.02657262887806, + "batch_question_ids_sha256": "8ec776449fb8174bd55b829feff0711ca000a9e09d903e741da6f91e2d21a136" + }, + { + "step": 49, + "phase": "full", + "loss": 0.16362383612431586, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 36.99050153326243, + "batch_question_ids_sha256": "a20669c6cee5e157b8b54729bb88921f4ff6df89b4dda467ec25abb10f0e78da" + }, + { + "step": 50, + "phase": "full", + "loss": 0.2581261422019452, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 37.71963926590979, + "batch_question_ids_sha256": "cab2a2f34731e8a959300614046a8b8fc4c926971dd5909a1a6e3de0ad3acb26", + "dev": { + "questions": 372, + "objective": "observed_outcome", + "eligible_questions": 372, + "target_ce": 0.5806565442146586, + "target_kl": 0.5806565442146586, + "target_tv": 0.27957613891343785, + "by_target_kind": { + "observed_outcome": { + "n": 372, + "ce": 0.5806565442146586, + "kl": 0.5806565442146586, + "tv": 0.27957613891343785 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 372, + "ce": 0.5856488371589073, + "kl": 0.31564624455137563, + "tv": 0.18229928321525732 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + { + "step": 51, + "phase": "full", + "loss": 0.5709785744547844, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 44.23742653615773, + "batch_question_ids_sha256": "46758e1c6996a3f1a932a11e8372539949f8ac8f4c0415b36f65c713e195096f" + }, + { + "step": 52, + "phase": "full", + "loss": 0.22201280656736344, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 44.957907737232745, + "batch_question_ids_sha256": "d87fcd81d136c657c13e49aa033db193bfbb121e4a721fa86c16aad4c0a7c39d" + }, + { + "step": 53, + "phase": "full", + "loss": 0.3642621487379074, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 45.691871273331344, + "batch_question_ids_sha256": "4fae217aef8823502570d85ef34334cdfe2994ba567eef62990a60ac5e3ede4a" + }, + { + "step": 54, + "phase": "full", + "loss": 0.9462117776274681, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 46.41959032230079, + "batch_question_ids_sha256": "edbcfec53ad977392c69172be6901bdb4eee1725bf8fd83582f64ac6320e77cd" + }, + { + "step": 55, + "phase": "full", + "loss": 0.618944376707077, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 47.144659521989524, + "batch_question_ids_sha256": "7af8e2cece43ae333d4687839255ee144e5968fc9aa063efa2ff6dd96c42d1b2" + }, + { + "step": 56, + "phase": "full", + "loss": 0.520059403963387, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 47.87055600807071, + "batch_question_ids_sha256": "6d579036d49f401569b6e29499a444725fe7a64de0907bc07bd389fd352e615b" + }, + { + "step": 57, + "phase": "full", + "loss": 0.292273685336113, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 48.57758554536849, + "batch_question_ids_sha256": "848946e51c61b43a984fc49047cb8c9f0fae9f28c6447d8b4ee2c74dafe198bc" + }, + { + "step": 58, + "phase": "full", + "loss": 0.47000307217240334, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 49.29666138906032, + "batch_question_ids_sha256": "7a207cb4e3ff6157802e3b560ec0b63a488a8ab0eabaa5bcdea45eeeb107d3da" + }, + { + "step": 59, + "phase": "full", + "loss": 0.4892919175326824, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 50.00856906082481, + "batch_question_ids_sha256": "6ed68210f517203266a035e4dc739a5e91ad1773fb37b2d954aabbc97ce44cc5" + }, + { + "step": 60, + "phase": "full", + "loss": 0.3904850035905838, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 50.76625431701541, + "batch_question_ids_sha256": "cbce464330f46f2f323539fb9d7962cede14e7074be588012e1dfa684881e6d4" + }, + { + "step": 61, + "phase": "full", + "loss": 0.3799959197640419, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 51.495480451732874, + "batch_question_ids_sha256": "7212e07d655e85ec9b791cd6cde65ec3e9ebfa5224d1854c76d9a8e98cc07be3" + }, + { + "step": 62, + "phase": "full", + "loss": 0.37555741146206856, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 52.22409507166594, + "batch_question_ids_sha256": "768e8960dc65d1f0cc4b13cc2c90dbfe8db2c685cceb3a8ed4f834e59a13a2f5" + }, + { + "step": 63, + "phase": "full", + "loss": 0.5196910500526428, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 52.95018903538585, + "batch_question_ids_sha256": "ce0e1bf7240822fc1f373178c20604e9dba0409aa63f2b51415a5f4db012ea78" + }, + { + "step": 64, + "phase": "full", + "loss": 0.5655737966299057, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 53.69126392249018, + "batch_question_ids_sha256": "6b8c351645894c06829c8dffb3045f3a3081c0fcb0b24c924b9576410c78d47c" + }, + { + "step": 65, + "phase": "full", + "loss": 0.2570880688726902, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 54.4230572860688, + "batch_question_ids_sha256": "06de39fb8b5de5ce866344dcb10447ed159378857789cc67daf23061f43e3e7f" + }, + { + "step": 66, + "phase": "full", + "loss": 0.4608404226601124, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 55.15514244977385, + "batch_question_ids_sha256": "1728e00f1c3d6bfc533f2891740148ba51b40dcb766709fed2a21437282f6ee1" + }, + { + "step": 67, + "phase": "full", + "loss": 0.2939162664115429, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 55.88841801136732, + "batch_question_ids_sha256": "26b40bee228ef411ff536ea5c58c9a2fe4790fcc250d1ec97719f5c2b6051983" + }, + { + "step": 68, + "phase": "full", + "loss": 0.7479757890105247, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 56.6233723340556, + "batch_question_ids_sha256": "ebcd0c069c1aad6db15af178bfb86b95232c842c724cae991202c127dddf4df4" + }, + { + "step": 69, + "phase": "full", + "loss": 0.5635014213621616, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 57.35782457515597, + "batch_question_ids_sha256": "021f795640c8ebe5b45f63378c3fb38ba2c849f5d70b747f653482f046de3fba" + }, + { + "step": 70, + "phase": "full", + "loss": 0.6172329038381577, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 58.09276952967048, + "batch_question_ids_sha256": "968367da65a754ce0d7eb72abb8cbb64b558c86aa9d8adda10e119723caa2bb9" + }, + { + "step": 71, + "phase": "full", + "loss": 0.3874517008662224, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 58.833030235022306, + "batch_question_ids_sha256": "5f2d81f647af85f4a48f30464c10a61c6a85b7d5d35f687d527ade870b0c1e7e" + }, + { + "step": 72, + "phase": "full", + "loss": 0.7938513308763504, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 59.57471597567201, + "batch_question_ids_sha256": "6780ef3f330f6cd45a377b3f3c501ecc3ef2ae65d08769f207f4a68c152e3dbe" + }, + { + "step": 73, + "phase": "full", + "loss": 0.3528990224003792, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 60.321714212186635, + "batch_question_ids_sha256": "d521c674f97a615228053a262b84518189f9580e2c73b32eee51bcef5ed76d29" + }, + { + "step": 74, + "phase": "full", + "loss": 0.4417751505970955, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 61.217427681200206, + "batch_question_ids_sha256": "6c26d6bf906bf556525d65a2709f2f59a7cee21e2f62c8bad8c7015fcde12f89" + }, + { + "step": 75, + "phase": "full", + "loss": 0.45803333818912506, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 61.9478549240157, + "batch_question_ids_sha256": "afabf643cddd0692428615c4aa81345acbbcdab1c52a5c9a48c815f9e2700755" + }, + { + "step": 76, + "phase": "full", + "loss": 0.5286638140678406, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 62.70219527184963, + "batch_question_ids_sha256": "ba8ce0da2f777532f12642355dcce1c4d29f8e0865010f26b4bf03c56d7723d1" + }, + { + "step": 77, + "phase": "full", + "loss": 0.4356435313820839, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 63.456245011650026, + "batch_question_ids_sha256": "9e0a4687654abac2570a41f8674ab95fcade2ac8e5e238d4661499dff7d58834" + }, + { + "step": 78, + "phase": "full", + "loss": 0.3878141790628433, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 64.19278421159834, + "batch_question_ids_sha256": "e36ae8ba0c5eaffb1a534d69421fb9118e4f45453e90b83cbbf5f29ef284cafc" + }, + { + "step": 79, + "phase": "full", + "loss": 0.4297530949115753, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 64.9118534643203, + "batch_question_ids_sha256": "171d14c1b2acfe9895a9353fe3dc68dcde6a998182a601a2746bf8fb10de20e3" + }, + { + "step": 80, + "phase": "full", + "loss": 0.3846726156771183, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 65.67086058389395, + "batch_question_ids_sha256": "5091eab21841dd9384edd2c719867375591caf9d084ba03c065878eb451a5674" + }, + { + "step": 81, + "phase": "full", + "loss": 0.6219327598810196, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 66.4101623352617, + "batch_question_ids_sha256": "f94cae69521603c0b60eb1cc6bb5370d2ea3d089f5b6bfd9a260b41a766ff974" + }, + { + "step": 82, + "phase": "full", + "loss": 0.38340673968195915, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 67.12694064807147, + "batch_question_ids_sha256": "f804d3278fd4930c24d09af7cc2e1fc70170c35cac45cbed9779d59f2d765e08" + }, + { + "step": 83, + "phase": "full", + "loss": 0.3208001106977463, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 67.86835037916899, + "batch_question_ids_sha256": "19b60c4f93e1ba7368b1f066c22c4ba32b8b66981562b2cbe2485da7433c07aa" + }, + { + "step": 84, + "phase": "full", + "loss": 0.39219067245721817, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 68.58466890268028, + "batch_question_ids_sha256": "ebdf0b28c90afc00f6691d1a09af18f9f66ebaf71377d0ba210fc4892869014c" + }, + { + "step": 85, + "phase": "full", + "loss": 0.3839174583554268, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 69.3296839715913, + "batch_question_ids_sha256": "4f04bf635d321e106f9856ba99a47fd6c2c033ae0450aa32c6dfe595a74ae4ed" + }, + { + "step": 86, + "phase": "full", + "loss": 0.278085982427001, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 70.06139855179936, + "batch_question_ids_sha256": "ba9c22de6939f9c021221fe5a43ff5f50b15a70b21ed4effe8d97caf79b049ca" + }, + { + "step": 87, + "phase": "full", + "loss": 0.3127078302204609, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 70.79180381260812, + "batch_question_ids_sha256": "c9f35e67c5374cdfd4ecee262a8723f0f688d0387f073072324c9d259d9f4323" + }, + { + "step": 88, + "phase": "full", + "loss": 0.5134549420326948, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 71.52020991500467, + "batch_question_ids_sha256": "64ab5437c436a5c934bb05696c9b73f14c9838d3b210eeeb1114438303808400" + }, + { + "step": 89, + "phase": "full", + "loss": 0.2528883069753647, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 72.25572444312274, + "batch_question_ids_sha256": "6ff06537c7c8217ebac7a64762ff5ee340f367a0f7652d25d5cd799c0d15cbac" + }, + { + "step": 90, + "phase": "full", + "loss": 0.3719199150800705, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 72.99558722041547, + "batch_question_ids_sha256": "8fd09d36031bf8ee878da3b38a0694c5e416636fcd74e603f03dd205a9a3ab33" + }, + { + "step": 91, + "phase": "full", + "loss": 0.3900993410497904, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 73.72918267734349, + "batch_question_ids_sha256": "322fb22dffd39fb986e5bfd20ac78a9963d2d5b01978598e3ea9ff285c9b09d3" + }, + { + "step": 92, + "phase": "full", + "loss": 0.3622107747942209, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 74.43911021761596, + "batch_question_ids_sha256": "d0417dd44fb2608169bd7f2ab17d57acd203be7f8242f3c2e89ad317009a2e56" + }, + { + "step": 93, + "phase": "full", + "loss": 0.37237807689234614, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 75.17962502967566, + "batch_question_ids_sha256": "57be67994833e0a92fac3ba6c7b6b1fe3b8383d53284136aad0b9a926f82aa1f" + }, + { + "step": 94, + "phase": "full", + "loss": 0.26953236758708954, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 75.93803230766207, + "batch_question_ids_sha256": "9c3cbfa9806ac2c618d0a1e3ca3ef41c91b3a86c8aa45a36071383124ee71262" + }, + { + "step": 95, + "phase": "full", + "loss": 0.40398708917200565, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 76.67349392175674, + "batch_question_ids_sha256": "100fb8785fb8c78e98988f555871d3d9a4ed2716c9bff93597e8a64de5e2db46" + }, + { + "step": 96, + "phase": "full", + "loss": 0.34487732127308846, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 77.40921703819185, + "batch_question_ids_sha256": "c781c81a212d6bc7153d4ecdfb49ff5f6570c3bde61d9f7babaa983e433d7348" + }, + { + "step": 97, + "phase": "full", + "loss": 0.3990098088979721, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 78.14464715123177, + "batch_question_ids_sha256": "8b19c5c7751e07c22469266bfa40bb1eaac4670fe203a8c3e8a289350956893c" + }, + { + "step": 98, + "phase": "full", + "loss": 0.4635494127869606, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 78.87295348849148, + "batch_question_ids_sha256": "043573f3a457c1f8cfbcefabfef11e430a1dd573e779dd7e19f7dc26a4a1ebad" + }, + { + "step": 99, + "phase": "full", + "loss": 0.3696139045059681, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 79.60388481710106, + "batch_question_ids_sha256": "933b002848b57b1ae58768d4d38da9823ca19226389dff7c41f6ba3e58ff14b6" + }, + { + "step": 100, + "phase": "full", + "loss": 0.26828806940466166, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 80.44768039882183, + "batch_question_ids_sha256": "2a2b48544fe4380e81979729c21f00410faff377b407c4afdee1bef7530a71af", + "dev": { + "questions": 372, + "objective": "observed_outcome", + "eligible_questions": 372, + "target_ce": 0.47277030621569244, + "target_kl": 0.47277030621569244, + "target_tv": 0.26819442928540077, + "by_target_kind": { + "observed_outcome": { + "n": 372, + "ce": 0.47277030621569244, + "kl": 0.47277030621569244, + "tv": 0.26819442928540077 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 372, + "ce": 0.4926936688108985, + "kl": 0.22269107620336712, + "tv": 0.1714566637835109 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + } +] diff --git a/variants/events_paired_seed17/backbone_config/config.json b/variants/events_paired_seed17/backbone_config/config.json new file mode 100644 index 0000000000000000000000000000000000000000..b815be77563f3eef88baf4b4c8ae59706e618ad2 --- /dev/null +++ b/variants/events_paired_seed17/backbone_config/config.json @@ -0,0 +1,63 @@ +{ + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "bos_token_id": 151643, + "dtype": "float32", + "eos_token_id": 151645, + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 40960, + "max_window_layers": 28, + "model_type": "qwen3", + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "pad_token_id": null, + "rms_norm_eps": 1e-06, + "rope_parameters": { + "rope_theta": 1000000, + "rope_type": "default" + }, + "sliding_window": null, + "tie_word_embeddings": true, + "transformers_version": "5.17.0", + "use_cache": false, + "use_sliding_window": false, + "vocab_size": 151936 +} diff --git a/variants/events_paired_seed17/best.safetensors b/variants/events_paired_seed17/best.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..c1ba7f5ae85c6d32fe4946e32b30631c1ddf4844 --- /dev/null +++ b/variants/events_paired_seed17/best.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbefcf829b9e36253ef46ab076f7d8fca00ffb33762e8358dbe780b3605bb777 +size 2385039280 diff --git a/variants/events_paired_seed17/config.json b/variants/events_paired_seed17/config.json new file mode 100644 index 0000000000000000000000000000000000000000..8a5c434c71812a218dad1fca077b279cf27f6dba --- /dev/null +++ b/variants/events_paired_seed17/config.json @@ -0,0 +1,69 @@ +{ + "input": "data/scaled_games_v4/events", + "output_dir": "runs/events_paired_seed17", + "model": "Qwen/Qwen3-0.6B", + "revision": "main", + "init_checkpoint": "/home/rwang/openjev_codex_20260917/runs/v3_teacher_coords_multi_seed17", + "objective": "observed_outcome", + "loss": "paired_brier_pg", + "reward_samples": 32, + "gradient_checkpointing": true, + "set_head": "attention", + "steps": 100, + "head_steps": 0, + "batch_questions": 16, + "microbatch_questions": 4, + "max_microbatch_tokens": 16384, + "eval_every": 50, + "max_length": 8192, + "seed": 17, + "backbone_lr": 2e-05, + "head_lr": 0.0002, + "head_warmup_lr": 0.001, + "precision": "bf16", + "disable_native_triton": true, + "validate_only": false, + "self_check": false, + "schema_version": "openjev-decision-pipeline-v1", + "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "initialization": "local DecisionModel warm start, fresh optimizer", + "data_sha256": { + "data/scaled_games_v4/events/train.jsonl": "4cebf698e4b71dc37d4d0e985b6577dd565ac370ac82b6f47f375aa6af69f8c9", + "data/scaled_games_v4/events/dev.jsonl": "0c352ee1064a5fb5d2aacad59cbea57afb897886d3d09e5ca108c866a5f893db", + "data/scaled_games_v4/events/calibration.jsonl": "4a8a8a2b81c3c46df6bb4c8690e97374dd0165cbef40566bb7f87c1f425d5562", + "data/scaled_games_v4/events/test.jsonl": "884bec71891668199925809d8708347f56f7bf06e8363e4a436347967b1423e3", + "data/scaled_games_v4/events/ood.jsonl": "3c00d845524634ca3b34559761cdaf0c52b7b7633d5517d1e7ee7f31697cea48" + }, + "deps": { + "torch": "2.14.0", + "transformers": "5.17.0", + "safetensors": "0.8.0" + }, + "gpu": "NVIDIA A100-SXM4-80GB", + "parameter_storage": "float32", + "forward_autocast": "bfloat16", + "parameter_count": 596250498, + "train_questions": 1124, + "all_train_questions": 1124, + "schema_counts": { + "records": 2356, + "questions_by_split": { + "train": 1124, + "dev": 372, + "calibration": 368, + "test": 364, + "ood": 128 + }, + "eligible_by_split_objective": { + "calibration/gold_distribution": 368, + "dev/gold_distribution": 372, + "ood/gold_distribution": 128, + "test/gold_distribution": 364, + "train/gold_distribution": 1124 + } + }, + "loss_description": "equal question weight; complete candidate normalization; sampled loss is a score-function surrogate", + "rlcd_candidate": "independent paired categorical proper-reward policy gradient", + "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection", + "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing" +} diff --git a/variants/events_paired_seed17/summary.json b/variants/events_paired_seed17/summary.json new file mode 100644 index 0000000000000000000000000000000000000000..ad8e558961931f83ebed23528519682928a14ecb --- /dev/null +++ b/variants/events_paired_seed17/summary.json @@ -0,0 +1,136 @@ +{ + "best_step": 100, + "best_dev_target_ce": 0.45392147839038, + "selected_on": "dev target CE", + "objective": "observed_outcome", + "metrics_by_split": { + "dev": { + "questions": 372, + "objective": "observed_outcome", + "eligible_questions": 372, + "target_ce": 0.45392147839038, + "target_kl": 0.45392147839038, + "target_tv": 0.2908940469690758, + "by_target_kind": { + "observed_outcome": { + "n": 372, + "ce": 0.45392147839038, + "kl": 0.45392147839038, + "tv": 0.2908940469690758 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 372, + "ce": 0.45846966947460877, + "kl": 0.18846707686707742, + "tv": 0.18377175030687837 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "calibration": { + "questions": 368, + "objective": "observed_outcome", + "eligible_questions": 368, + "target_ce": 0.4863960135417897, + "target_kl": 0.4863960135417897, + "target_tv": 0.3079795091580387, + "by_target_kind": { + "observed_outcome": { + "n": 368, + "ce": 0.4863960135417897, + "kl": 0.4863960135417897, + "tv": 0.3079795091580387 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 368, + "ce": 0.46718815035607203, + "kl": 0.1808918312753622, + "tv": 0.1788949192262398 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "test": { + "questions": 364, + "objective": "observed_outcome", + "eligible_questions": 364, + "target_ce": 0.42791067276663786, + "target_kl": 0.42791067276663786, + "target_tv": 0.2817070426047916, + "by_target_kind": { + "observed_outcome": { + "n": 364, + "ce": 0.42791067276663786, + "kl": 0.42791067276663786, + "tv": 0.2817070426047916 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 364, + "ce": 0.4476948609498736, + "kl": 0.18692406062375522, + "tv": 0.18073862322396844 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "ood": { + "questions": 128, + "objective": "observed_outcome", + "eligible_questions": 128, + "target_ce": 0.3296768191800103, + "target_kl": 0.3296768191800103, + "target_tv": 0.24400384461107302, + "by_target_kind": { + "observed_outcome": { + "n": 128, + "ce": 0.3296768191800103, + "kl": 0.3296768191800103, + "tv": 0.24400384461107302 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 128, + "ce": 0.3458871208971331, + "kl": 0.1123998992605057, + "tv": 0.139747922104369 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + "temperature": 1.0, + "temperature_fitted": false, + "training_seconds": 100.81704722624272, + "max_gpu_allocated_gb": 11.964137984 +} diff --git a/variants/events_paired_seed17/target_audit.json b/variants/events_paired_seed17/target_audit.json new file mode 100644 index 0000000000000000000000000000000000000000..af673c3c3971ef76edb29b4a5af29ed643785342 --- /dev/null +++ b/variants/events_paired_seed17/target_audit.json @@ -0,0 +1,40054 @@ +[ + { + "id": "event_v4:4aecd3d8d87778bab6faae73:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a5cbe3a96ff9859f1afbdb0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1fbbe349381af7a541cceb35:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4db212f04284ad6caa8a4724:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c1e2ae6e3efa6052c93e30c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5378d5f62c640d6e0fe36b0b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f24e884a3b8605a927a55f83:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f458277aaf7786d8fe20847e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c644d0d083026cec6c153e9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ccbf6b8f9a588cc5089fa918:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93036e72f8c9f2d9ea0f0b02:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34792163cd9a8f97c1f4603b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69fddcedb520a75892669dbc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb610da24b02cd384933ddff:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdb45bd714f3c6fff339413f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6b7a7d7738d1c2a9874105d7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b083ef1931c7071e74652dc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fe27b46da3da7e169a4927d0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:196a1da34c2b563ecb152fe6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ec30ea0adfeee4fc39690d83:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9fa06a502ac89600ed238558:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35674713480dd730da6c2556:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b247183873660302da213d06:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91f994fed8c8d4008fafd666:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:46a6e10623052a887d1c2205:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c986fc04c6321670f2dcf9d7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c32b797ced7f503623b6407:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c8f73319f7da5359e93be37:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:168f30e6158735dcfa4fbe3c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e69848dda7a61a77ec326915:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9583dbb10b63349ac8fbd042:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e4ecc9a3d8526c6c4c36d32:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:778d5649938871adf7e3fe56:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01b9d89efa545728f8e9ff91:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a221ee0f166a7fe69dce2d60:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8683abff8c9fda846ef54b5b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9d8e40c9559f1335400bb1c6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e345dd2850eaaca29fad266b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c83318f0abed62232f75b3b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ce9228ec2e6d3dcecd19718:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3156d2a5861f363f4847de3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9bfe659d2661f885e63407fb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:19cd809568aa6c69380928dd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9e9b833eedec6d2516b563f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80ba1b17757572d541430961:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e84d57de9d328c6af74168a7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aab346f5a68e8b0d9cb8efc6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c9bfa2a376a324b78b76000:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bbc9efd11d627b4cf460c754:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5836727011760fc93a5ce716:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0afdffd76f8f8d912a2a9810:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f049876cd8d7bc1c1700c97:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25dc1b4da5ac8363f6d93624:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fd21575f55b2a921f658d041:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a95f0f6eab2d1da5dd6ecabe:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa7005c6ed1f9d7125195daf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6d76c5dbc7687767aabe29ce:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:840c07180c531dee4b614733:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:496c9ca3430d9a20c6f7c687:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e3afb50a1a36f19d4ae0b24:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:930568d553de51b6a5d759c6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8069fef34727ee2a1902a848:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5245ced54ca26c5fadec7646:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a4a10738fe569740dd0f7bd5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a6fb81b728f8dfa47a7ace4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93719e4384a5461e7a7219bc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12482d3af71eb59bebc3ea10:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ebf4a770693beb8e19ebae4c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a793ea6fc8b9e5d41771c5f9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:17a051caafaf5a77999f760a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f14ba7d8996c6f860ead6fc1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c79f246342ca3183f470e4c9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcf61590f5cf4749aa1280c4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7928a43d654582bcc499f60a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:510689d0c457bfd6b4ad0a50:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47177f54b2ff163c6def1b2c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ff7f129a6f7732d7e4795e4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a807079a25b56c67ac66ccb2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b0436e119d314b5bf86c248:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdfd7b16eb8dff830b5547f8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df1da3e45292136ec837892a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:758a0dbee8da7e152982df83:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8d56339aea5d5a95a5b665f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a88a6a2f1f0e68bc15a24c34:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0147d79e60563e071a1dd396:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:086cdaf3d5ab2ea94df598dd:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c718fdb80210fd11b08e329:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab041a7207d659c5fef22d4a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96deeefaa3365cda24ea483d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b74f223bff9f3d72331b031:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ff604244fb4de2bd94f9aca:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8300f14e76d85672a8c70669:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:666ffe69714f5e4b0d8a1dd1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1674014088efcac3fac914c2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e87a568ac2f2e78b210e821f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:254f7551bbfb139006266b6e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83cd93517670fd958ecae5ee:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:634d8cd2256448b12a2ebc37:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:081c18fbe82a3702fe712cb6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:860afa1ef370b2e6f70bf40b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:182efededded37823893473f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3d7e5dd2445849704958bf60:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3094e2d4d0cb34b5bf1d209d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:901067852a870e821b821b6f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d3a0a64ff3146b2c4360e4b4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e856c6d0068ced00cf2a375:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98e210f7e13494da7390809f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29f555194a95a09dfaa16eff:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c33b69ffd0e5d993162a45a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c37e3a826fefc4b83f1afc5d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7aff4ff3d50892441a8f2162:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc2d86938d5220b4ec91b2ef:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7b7c87b3eceb318ec6b95df5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a46694e1cf315b823b32b1eb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1424aa2a069f7812473505b8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4532e9616b169212296aed5e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ef476d69e35e007ec2ed8bb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b295a533b5c8c0ec6a94727:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a76c9ad5103bb6f3d9659029:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a3c6b25ad42df48f7a7978e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a05fcaa4caa3a3115129fb2e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4dc051c0a01fa2169c8b9244:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41b15a05073f28387831b691:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0dc938f084a2494a91529d71:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee54ccca139e1cf56c89a462:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a54e3adc5cf2a695bd83f471:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36aa1f5ef31c1178a822b653:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a6f3af0186a53a52c1bb58ef:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e6c7466ac0e8dd4278a1cb8e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea5ba6ec793a1e9bd609b765:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93cadcde1301139898ef49e7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d09ef3a5bff29862f750ced9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddac902fe22413b8b46805f8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d973b63f3df4fc07d52ed79:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cae384c4a93e87b360c116de:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6cca008fadc530233719175:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:538249383f3cd0a4e1dd2e1d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0edb8696013ec30415cd3deb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:554d0d960937239f75a5af2d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:52b8b97eb0c8779ef5da3933:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2cd0d4d1a69ce0bc5b4c649f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2898792139f17a9c4384afdb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d3e082aa226d2f5e24dfe45:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0214fa91d48fa1b6a1f894dc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b62b892a97c579a76b0f53d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5203bbe1d03bd01e2ca478b8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cba64bb97c7ab78c0fe29324:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb8fe05a7b0c62029f2e6ee2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d7193659271488065d5003a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba98d4315a4cbedd4e2a4a59:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d255c6ef457e46563b4b57ba:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6af87b747f689c9b4888711:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67d1f85b1befe76def8fee51:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7d9482c148cf9b95ebc1caa:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1d3310aa8547411a90a3bb7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e6b9dfd3c07d3ffa2205fd6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1827722f96e17db2bd6fcd84:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ea73ea02edda40bc54c044c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b4920c9fd2938836826284f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e027fbed3245e5b946c91da:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01f48f85232115f71836ae92:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01ee9a28d0f9469c007b5cf5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac5276e3d4831247884eb8a5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a95af135a3716d9969ef0e21:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b2888c81393ceb3d87874a4f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62263c53b73cf9c7e1c14975:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9569d605fdc66bb6c76a935:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85c6a22076b02b74573a4a22:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:616b96b76ad67d4c9deafeb5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:00d0e1e49c64591e5f0b158b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f709122ceeccdf38ad920fb0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85a36ad246c92d82ed439aa8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a8173dea9319f4acbdcd0f8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c93ed08d1f079a56d20cb73:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89b8fc1389543cc85a9282cb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e88d9a95e134302a29ec302:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd8e11eb57380f46a4533532:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3382186e120c20e414cec7ae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:af836278eb0dae03ebe1653f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d801e9df52ed90515626e65e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2afaba5e89d1bf4e970cdb9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e46b02ee09f14a9af0ee7bc0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22b1e3da175260251826164e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7dc426e15a7fa672e29109a4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:002b433b16ecc07c398aa3a0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c752e38ffab6e0a3117b779:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e1f79395cebd5fb8251a78b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b0f6359c5359f4b9d09a0cd0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1f199a87641482c04355483:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a9c472da2ec4b16b5e252ad6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30e895a88955e8e1741c3cce:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a4a29c08ad1ef7361d6066c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9299075c97ddb34754fbf99b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61046bb8aa61abef3bf5c069:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e812c569fbd2ece72b6ebe12:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a1ca87d4133b30b1b878c04:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c6877dfd63449d3097ef947:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4a07f731d746b4a32a673d45:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4a54319b85b667f85c1f87b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3a8aa14cf117a3c0edde5594:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca30abbc0d1a82277a7df57f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bace6e281ab7070db1cc530f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9bb6c22f14d1dbd762946aa2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7c4f642e1f7ed06decd0ea0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40bdb7a9c10dff12db44a3cb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b117b65e1f5ae573455bc7a9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a54baff6329b66e6db767bf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5533620bcffa09321a5d0f83:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0b73a17456822038153a599:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0d6c54e96ac000bf11cfbd0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d13966c3efe0950ebbb60c0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40ec2c07a3cab13452482734:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e2f8cd7ab90d301bb1b973c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85bb2b2655da536beaa4df6b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7d124aa6d2ffca932231b33:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4fb18e94b72edbadd4726e49:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:199a56e347100553975c9f2f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c36fdd7d4b562bc197282912:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8691b9f59d2f19d6fa25801:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:833b703763757af5e4b68b35:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:52720ad7a21f333c077b2110:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d932e48b4fddfa51788a6355:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:252d4d472631153cef78cc8e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0168631aceec340c766a850:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b80331706205a42292cf1af9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:248fb4f0ee7f5be24f49a69b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b389bc2b0a17e007ec63866a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41b2f786c19ea8833f6f57e0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54e3862145e26d2759ba5ad2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ad5a0c8f22b27c96b4b7a111:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7b9677016d81a9a7b746276:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e78bb8f32a3125554b5eba2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d01821e72b6a2a46cf93409:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e983594b520fb7548ca6a6e5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81ad6fb442bdffac10981404:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50bc7c88108132050d595467:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:221e3b56cd290837c7f03085:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b759eabb218117d5b4b1afed:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:549f1b7f97212fa1f78a3c5b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2493d89ac0d2715ed9bfda35:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30bf9428df242ea371274770:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d8322cb4fb052fd1ff83efbf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c4fe6f890052ee219851997:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:04724d4d373009c3c1df0eab:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a3cd9d92de85f633a5dec4ce:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d8481054b13cb02eb33a1e9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e0898e9b40f457a17d9026a7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b53d5d0ac873cde248198428:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b426a808a043cb0bd822ca3f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f70a11c264b317566cd767a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:43f61961b03ca186cd8212b0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2477ca30a8a97a0e4d25416e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d914d7d82fb5354aa46d8b00:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fc045fed1f8537f357cf453:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80243610db5da348b2a7def2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcc4e6fbc8c6ed9fdf00212c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:987b74ea0bf071b910f7e810:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87a1cfef0ca559c4f66887a4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c17fc3799d7c063b5c4bfb3f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4a30c805073de74309d662d5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a41ab14b6ab104574384653:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d01cc3d9771456d0369920b5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb9c8eeff8bbce90cd863a26:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb3d2173c7b2328ec6752939:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0652e8b64a4ed5ba0d647126:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf59db09948fe96a90130fdf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c3adceac18a6d6e8c481e1b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba9a4a75a947f371d1f5a481:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9cb3ba48fc6d5384eb585f7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:292d86322c0c34e7f5c7f30e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aeefccbec51a478f2223b626:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd71354748063f8d30e80b4e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fbaf2c0c62622cc533b401b1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:239ef25f5d857b47d242d6f9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98252467d7707f75aabce2a2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4092f7aa8f3083cfde4c58a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68eb85b4f7c251c9d3263f6f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7361c28dd204b7690a52e4dd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:236c20ffd93d26e8a865b73c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f3fcb67d58c6cf28a08190b4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b2add6850aeeb55fafb5066:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e439ad1b03f92a35d30381b8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:74db8b8082d0ec7061b0dcdb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e595c3863ba7c0335d80f775:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1476189f52ae003116af8f2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:104463a9e824c39ca1c78ded:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:679deb7673d4ab6a1ba5806b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fbc6594a3fbc5d7ddf80247f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2668a50afb39f45a5708eb79:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d578a3b877bda7b37b471446:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:938ea4a3f43edfe150282098:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f694ff937dff30da47693fd3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:990d3cba4db1fc53302fec38:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc42a7f86d3a9a6eb121adab:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bfee33f042bb50631e54845:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:789715bdf6ab3a2df5417bbf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02e0e060d800c08a7feb46a0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9432f491c7a3010f7c860661:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e691b22b64708e4cad8a4a6e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9350dc6743fbf0e2f2189507:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61a94f388e9f8f2dbf14d11c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ed16481b7895086a6ae0910:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f26146bb5865ce7081895d45:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25ceb291cd0bf06ba1da25b9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e1a74bd5e20eb6c116f78865:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3550014400bc31035676c567:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6369c20a70752ff7ccd6688e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5e6366bb9d2e3f6488841909:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d89fbc2a6938a5591c102d0e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:807bd7bb545f3363c43bc83a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f5c0e4ae9fa09e328c450e8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81dc5e532c6f1e076f70eb89:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7bd683df9b5f1f554ccac45d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e97dd3e41a15710dc9dea1f8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:214b3693143e52555cf9d3f8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3f1debc5bf500f6a54ed3afc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab0b47508e943da5fc0bc22c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6b16b1e88101ab6ed3cbebbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4274505c3a6b1ea2236ea6b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c10e2d63cf9593db22f07112:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7615413c0accc18c251c8e66:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5b296bf9fbf65a3f5eb2335:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12de4df791b5835d3510a9f2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:605cbf99ca463999d7cc4493:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0df1da747fa694bcaa207045:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cbd2aeeb33bfca228cdb7f9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddb4b15cdce7f68834fc04c9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:71044f235933988434f26106:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e1b42e7bfeaae483aded5f2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1932357263b412187a3089c2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8da659044d73f614b7ea22e9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41b49028c2c8f400c9467e36:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:882efce7e9d263e6984460d6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcfd6d606d9043e8ce82e88d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d55f1182f70611ed3875206a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96f2217a5354b46068eace99:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f7fac7cd2de3945d1deb891:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:002a53a5aceda2ac2ce52bde:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1fd73bed926baca20befaaa0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c5196a2731657dd74eb325b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e85d3f46b855dad1c5f1ddb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:255697241911fb3c271cef5c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4048f2e42e8eb20762ccdd6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:138733578bc678ff32537417:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bf24a4008d41f915e1fe7f9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ceca5ba33881a0df60c9fefd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:99ca534f1179e15e7d58b376:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d68e4967f66d33dbd27f5a0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb2146f6dbbb9fd75f85953f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ab35fea550164a919ea8020:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0701685f4b125ab98b01f43e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:197a8ff15e4efdacf1d9d8d4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:570a426861edde554174aa2d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7ec7d81347630c4503285b9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:797e6d858c34074a543dccce:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b995f52aa70685be6dad342:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e22e5f2f2f18e13030c01be8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4d777f612ecf72fc257f59e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98a69fe82363f91bc871aa9f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e15a945584411fa1b32245fc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0505d0c90f7e5a0ef9db8097:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d2fdf33219471e56a1ce28a1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9cf1ec307abe4bf0da3ec2f1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a50eea7365b428382fcc3f0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68ec8b82a6b56d72d79afd9f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 289, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68964972963ceaa0251fd85c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 289, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6e2d754f6e8a30dc996b714:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 289, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:79e1eb53a8ad8b73f07e96f2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 289, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba7a4d97d9c5cd5a91633d7d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c01989785a6ed6406aba33c4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:349a5dcd9f4878e6b189e285:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50b714be855f405803e90240:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87b2593405d421d6b7a28dc4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c199d672a69e2d7636cb026:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8280d12ab7784126b42b0681:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:721054860789951044116121:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 288, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cefcae3de562b00d42c08dc4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30daabfe807e170a81e6eaaf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8905f1c83713bbda13b0616a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:358db027e477b84d72e96a03:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0822effc0b54089eacbdff44:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0df480a355860888d0a5602b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6135141e7e7b1db72833bffa:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2ae4acb563ca7fea8469e8d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0669a64365763d5c2f696236:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93db94cf414a8f414c8d615f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8dc21220b05303caded38e7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3423add5646023034abf055:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed250b8345eb1d699b876343:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11af4b1978a1f0af8421b399:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61b84dd7e63312b7d348bcb7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b43548aca5107e43226b56d5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34ee612afad1e99c2eb710ae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:108910315ccec0a3e244a7e9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac8cd90feabb6eec95839d72:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb6d76511fe15121ee9832fd:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92ff30de0217490919c52f65:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06a99381abaa185a6a9fd049:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4a247b41be9e261216e342a0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a774bf3a92f5935173d3d2ad:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30bca1ff0db749828f0cdcb6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdea874c3af71af330e8dd19:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8d874354308484a8eae3f0c3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0dc3f71b2a4699ddefe5d14:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ce616f50c50ea8f50311521:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e57fe3476a705cca6726fb0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdfa2566ffdf094b245294ef:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc273089829f9de3e5112353:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:975ab57af5710e33e2b59472:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d7dffaf2c214fe692bde8fe:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:17e3d8f8e49028635017cd03:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9fd92e91c1dff41dc113b20e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:302d27a5ea6a934d5ac42b00:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0dc4544416ed815d7aebee50:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2243504d61f3609cb6e5a783:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69c5810b4917634567135eae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92ec2c46cf996dc7f34871a5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68e607d993a088bae18435ef:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:215140a4cda26a73503d3a83:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f854dfabf546b974d504566:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea3b5cc1ea8122b10f5f86b2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1350d478ab3b32c8b9f6b2d4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:550880c731ebb9d5c09ec24d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ea00ebe76fa3029ac6c7fb7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:476cc169236574a7988d1204:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a55e8124c6fa131a565d4b0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:303d0afd60f97e56aff794a0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:14435878a8d37bbdcd2dd5d6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a5dc91b79c9a2a823af6c019:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:60fdb9d681067a1a91f0b72e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39cdec65c8851ecd19231221:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1e923c9fcc64fece7829151:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d725ba1fdce0dc5b09a394ca:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef90e306cbfce9424175126d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f6754250ce60629b2509439:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ece0427d8d7392715d70ec5c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b2901b20d25de32fd1f255a4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9563b40907302e25d13773b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:57fac8334503978da5a90a40:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:347d698cf044ca9da0e619fc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 276, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:53e1ed1fe556a8e1ca5c2c4e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01e0e74ed27a6dee36461958:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c8fd4f874d575a1de9108bf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d740012301484b159c0a0dbc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6c7874bfad93f70a1ead4e6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:030b79e3eaa3583b3ab8d5b2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d6540ee67a566ec60d881d48:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c85ba39c8672f608029f5774:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee2ce711eb4c97ce04e46bab:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8f7a4879de75203775386680:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18d809d6ec9834b79dd8a863:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:64be282f7f43715069cf8f74:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3a7ecb6565f50af24257dd6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb0db769fea9db7eeb4fee99:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:710df1cd8f2b3a69905f05b8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f370e94562544c6ef17f2dbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a00e5d8f9b4a206c81833d8f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3356a702168cbcdc97cda7bc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12e939a19ea31dc542251c43:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c536e43ba623c319d126fff:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:14c6f58596354ecb87920614:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2103a5c5f4a45b95e0d6194b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4f9de9d504d227b5a994dcc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02e2cde5d8ed423d57bd5ac2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:908b506f0688ec79e57590e3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38994ff5042b36b0c1c55692:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b0f0c71ffd2edf8da980852c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:046c8c58d982cc8d5a0492c4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4e0166391c4975325e16209:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85a6940d1acda164739ea4a6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:56eb58aaba3e308c1417b500:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c1ce6f064ad5b498c644532:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d68e01be9b4c7ee47e12ea8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba040a21090fb09f79e69665:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9806fe49684837640b12d62c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c1ddd083656018f58c197ae:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5c1e390d768827716f5df84:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:79649dbab1f24cf15a3ec4b8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30230f7ea8af16bb21d90e88:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1831705374f85e0a6e15efd4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5dc9447d071b881704cb146a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:118905145afd251a26d29e6d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6e96af1995885c22eeabc35:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f478c7ffc40d4f4a0f8a2a61:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4fafa439945d3eeb7c614c61:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:796e4152d808101e800aedaf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3610a9e1816e92532cc72c29:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e1d84d5ed67202e645eb63f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89c2f4199a905249e1082801:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:55febbbe1c3cf90506d136ee:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bc9947dd9ce76fa7a77e0671:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba3a43cad6fd0c928b4ea445:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 269, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7356066daa7cc23f0e0badac:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b575f801c426fbe0ab34b1e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d7036fb11be505f0a9b3a417:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6923d5f85c7161e63d26817:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8991a01dac500c3e224d55f1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e2b4f6fe5f3a3c9d0f7880d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cb61a080e4ce926cf069537:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6193292593bc19879ab8d578:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c234b0c1046e975e8fc5e09:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef728f9a49ee671e899af858:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e24f353136e07d3be10cb3f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91f7cf44076c6002bbe8bf3d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83548f2cc37415c0eff38d44:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:109ea27a31551beee79f4482:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e504926c21cdc72076f9db5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2359ed68991dc93c9796f8b2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:72173ff70c74944a07b4ce73:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a7c9684b7720da082c99716:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bb0aec73c3ebeffe25c7e9e9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:110e7c0c1c03f20e82096244:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fff52480c3aa521aeb29bf51:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91f323f74c229b1247c3b9b9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d1f435cd245a901f0684e31:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:56f96a9f41e776cbef2774c1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29ff8183f5bb30345cab4390:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68ae6bf18b181373f8d6d4e1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0c3f4339abf8944e23e5215:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:717557dd783541f595805baf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee59c6ae58afff53499628df:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83da5cf3d1b1ddf8fa763b5b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5c92a5701eb6668bc1f8fa9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e469a62013803fdf94f1f59:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c2a82df1e5467aa7dc6660ba:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:26687c4ad73cc5184df26a15:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fdb332d5f8e4a0d8f459118:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0990419723e11fcba0affcd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca629ed704a66b1357bac779:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcdf7e375cf726310e726038:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5399697cd98f72753ff4184:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ecefb3605c03f41a999129e3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:845e97f19a29261cce2314ed:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29f234ad5aeef176f1e44927:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aea75938c86c8105d0feda0f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbec129f76b8c83c86381f95:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b348cadf48a3d03624ea0ba5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4035835d41d16c3a1fa8b1e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fece30cc4adc4871484d458e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fe00df43f4fc0eb779df708c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02aa23c6f283014c1abe9676:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0fe3e35ecdb9a7802759f3f9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:32d2a54b10ddd91944c573a0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a662e8b20507e285a191bdc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8ab283396653bec8d845a1bd:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c472a5c6051d07553cfec417:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47b61fd930b156045842a577:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa5b804e708259e0c04ba99a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:90c51dbb35adcc95e4ab55df:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ba46f215c4d34e0853b6ded:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b2be4263a3dc775e51b23c0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5427484a4262ff3c7fbf9f79:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5d3f0b661708700e8809d61:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ac884599d6dff56fa18f63b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e696748406acd478daad6de5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f7ff868b5c88c0cc528c43e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3514d35e6a4c8c67dabd6266:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ae1c08f689bd686984af0b7f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23fbaf66af2fb9a5c43ac00c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cba82a24afb9a6c6230917da:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:42d501bd5d2639615f1bb65d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1a5c14dc584cdc571cb6207:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8f4bffd9564951db12676035:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f1785f4091164d7d3300e26:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:420d69c51d42af2700fdb212:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b46c387df42006311cd59d82:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bac467952578b22f282d77b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac815c6135999e5968b9143a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bbae94c54a2adedee667f34:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e195100b5bf4cdf9f9dce4d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f58ae09b633370a388beca57:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23e073489564cbc08d4bacff:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a9fa261051646e4c425f6bb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c2ec9a05e0056357c4a8eedb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f29c436c6df6cad38deb3b9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:09019c96afc17c3bf63c15f9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f667b9249bd4af8ee7f52b5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba6d482786393b7dfe37c6cf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2be7856c2e3c093585329281:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1fa20e1404323924dd8b65a7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92a812459196dfd57fd7c593:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d2145a80946516e84732608:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3313bc5d19b525468ec7bfbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cec6d95e463244ba34935e60:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16ae671da86347a735e04d27:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15d0937b9eb7bb64031ec594:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:48a65de2cc08cfe9b53b76c3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb634b1bb36bad706a7084b7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:babc582a0929c2d7c07c1c17:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc30f2c0273cbc1bb0be8fdd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0bb4b76303c187ae6a16176e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e4375c60569dfb66fb42573:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:af85993cc9a07a6e132ae3c5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e328f9c46b379e4a9b37d08:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e7e58cbf8db050114101723:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:177265c717257ef23e3d69f1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4250b698d7b9c10b232469b7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9dc149656015f6baf4f46219:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a726bb28414368e00ea7c716:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa1457e17b5514610771b43c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06652b5bc00184010dcb89ae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ac6b2b50471d709b003454d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b2293aeb8e3eaaeb1d16d0b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b0ac2acbf820ac72b0cd6487:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb37e69ed7dcdc938d2d52a5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67d5b01b97a26c9573f4d52a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a04687efc794322704ed778f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66bcbb7e600cdcaf38db10d8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3593695e6fa4bf47bf7eb60c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54dedf522236b05f475bf13a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:792aa4e64f9a2e723059bc0c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0ddf3066fe7a07c046f2376:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a3ed1939a8edca581359f60a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:27327bad4c5f81287d5139cb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f8cbd20dcf4f9ddc5c51d646:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:607aefb0c7d22d7bbcf0ac33:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0fb010b04e5ab00b94f9c722:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cb5cdd9c1b9dd9fcaa8e0c1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8857c7cc49ca8c72df0dc226:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f2dec7dadfa71e1ef0479e66:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0acc8b69146025cd4934c70:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:57f0093b52ec2b56430a80f6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0593b127ed99d3616c1584ee:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4cb5f335687cd8b201f6d4a6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:97ebe26d7497d43e821b8e28:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9cee84ca5c088d1b37392d15:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c4d22aadb729cb8b826be92:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd787b306f5b5f04cab14c66:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:94688c378413f969e8701a8f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8dc99806488ab6384e7aabce:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a3bb296ed9c28ff8b9a5d743:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2af1038af5283ce6600f3d2a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 272, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5738c688f78baf9bc6dddcde:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8f9796158617bd8c568a93d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4355fc17714a16342ec2b97:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9af375fccf968cba465ad0f8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:86dee5796e3794236a4ab2fa:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4cb9a623c1921ae41eb6e0d0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6b8c3864a7d66bf960feb6f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96e2c9083ebd21ea07e08c1c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b1ec99a0da91ebd52a4bc72:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:97d1a8e484488d2a6ed47ef9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c73dfff2e86789c2afb78b9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5747d09e72c2bdd2c192de3e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b219c24196b6b47e8990a71:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ce736a0e9a327a1e3ab28171:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df3045a4d9eb07f99fca7e99:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75c32f6c1b3f14c4bdca3a67:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0090f026f9e8985f16cef679:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:874fffbc6e8b724a715a1443:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7b8ec24d53082094ba6cfb7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2cb5e70aaa43cffa82c7e399:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c17e9d4a789e6976eb1205ed:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:17a10837f8c166221f9234fe:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f7df1957ce58aa1af8ec6295:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66f14148c22c4c918d32a583:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:068bcb7a7d9b80b0211ea29d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4567f0f02ac21971af6c2067:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5970165308c46d21f689248:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b820c12089d45b425a2f143f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8db344327e8832b431640f29:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30e5aacf3967d75bce836157:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2aacbffc4db777764d114fae:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9791481a2c8586a8b9359f89:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:150e2cd536a5bc1059f65dc1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:71034de81675fc16e59ca900:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:432c9d646994c0bc4a45c30a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9fc0ffc853df57b1434b3f93:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1cec312461e179bfe87fc6e9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc228262685c9af863e04f40:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0bca2d66187e300d825b0e2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:185cc5fb91e6085cf42bf0fc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ebbf8f64e89c61e0f31df873:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:256b930168cae56bd6137a15:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7ab006da8bd3d6f8f094feb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1dc1fb1bb2ab20d549e1b067:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:801cf4e3380a9253f7340871:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c8179a22901be601d02a0bfb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d22c006c056739961a9fcc4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ba08d18374086eb044a9a59:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1ae888b75b45fc321ae31b8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e182b61b80cdcd82a8afbcf3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f61c18f607c68048efc76767:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:760691b4a6b553e8e6753666:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5bc6da415ac3e3139c77604f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4f74c2cd88ba1556d014cc3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b362adce01db8b4de0df49f2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d91ef6762a4834e41e226d7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:da4878ece31bb24a97167d65:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f90237ebe0a2a0b99bda2998:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cbed873db6aab5eeace3694c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d56ef7f829097942226c38d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 275, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d56a2f30696951e16e25a68:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22d3b43e71ab03e8c809e636:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ce6114047f6b01f950d2c7c8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c68b5115e21bbffda6349ba:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 273, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d75093dc65ef08846569126a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:af9403a01d459574d8e44d99:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:204c301980a3c082bedfe1b4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96bedbaf1dc4a9c585bcfdec:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8cea37a22ec1d8811e1316ce:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e234ee54eaf6e14a7a3f023:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf8cac29d1b7416ff9a898b0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2fa298c5736a119c0f39ff7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4df98939ec82fbabd65766f7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30c9c386bd32f1dad5ba2fbe:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18eb3dba14ceb391c588078c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fae9510eab9a0c6118233776:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a45132adb7911e7510ad744:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c0f458485c6a396c24f942f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:20b549d9a6af728fe092ca68:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f9b6fa2af2d327776b882ae:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b057646709e1db1debfd35d8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69cc59bd8e24338a96430932:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2307ad191e5cf6f9c229a277:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a08b246a5ddaddabb4b86f4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:799ded37832184f9b0ee3168:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcb0dd55df41f2ba318a1bb3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b952616a0a1f108d33bd9eb7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23e2df6c0d3866801ecb2c45:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3dcb436a16a6492943a2fe42:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f1af8e03283abc44dee7859:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:046c78af4f31b3e1ab065ce7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:708587746543a9710eac6c48:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 270, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d22aa8347ce5fb1a6dfabfd4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a7fa96d2e7cadf4a5f3b6b5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6637b24f2a43ddb345c43f71:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80890ddbc902f2364a603ec2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:60fcb58976b841328ff5a9bf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:229a9b464754862682004324:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddf5a73e9542bb34db9abbf3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6278dbc5ca5e99ab00b0144:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a300c51301010ff18ce3cb0f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bb87e2d44e695ba430954bbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5b9410e7bc5fcfd47057e921:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:185eb6d6fef5f701433b1220:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2dacec95a2422de285a8dbe7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3d31488468fcf83a2c803e4c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:76cd3dadfb8a076af5a66d49:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5ba5569d43760dc324e2833a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b714492a3f71948f1882a11:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92cf5eaf90736c9e783955ff:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62e0f1370c4c1b99cdc6e5fb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:26b949ebe9b83c3aa170003e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66d868a7bc10477729ef2a16:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 595, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:618fae9872bb340930b47b13:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 595, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c7cf7b900b8f387ab2ba4ca:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 595, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7566f28d2f41ae8fda8908da:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 595, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:049d0bc4ffad5871cc6c21e5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e85148c1ca8064cb491e1ac:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed45a43e07ea52f3474f3c6d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa72e2e3de57338165467204:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:496aebb810c5df7c3fb73488:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e32308b867c8e6087e133433:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e984765547431adc0fe178cc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e4274c9d2832ec90c5d381a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 596, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3389c4d6994e16bac4135b4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 597, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef0fb44a5c127f5022120f02:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 597, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1bd52436398b2ea1bdeb6a0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 597, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b30fd38a909a162fa7d896bb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 597, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0dc053390957a7149eb3d08:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3b41c608f1490ca652273ec:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8dbd0e73903b75845c88813c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddc48f37fc6e4f36738bc86e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c9592fcc0e2bd4b29222061:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a57cff3bef65fd2e9fbf58c3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c2f66b1dde9743bef20b0266:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:337b5e5fdcc4fab7d3c53ae5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3f605aed9542453a8aa3b21a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b83faf6baf4766897e6521ee:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df78fdf4f7e75de43738ad2e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f166af9dfb76b10ebf7f879b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a78ff1ab2f9b43ecfc7ff3b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ce7589b9818b5720cf2933c8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83345349bb580680ce39a90f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5c5d234c0bf971c21b140ac:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab4e71d67dba088cb00b6bc8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 545, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:200a325d944d71b733e57056:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 545, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a7a42b1c7fec28886a15d87:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 545, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9e494705f31d7673ac4f282:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 545, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:739fe045f64a353251f53255:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb79fae11d1e445c1a56ef56:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d7fbc6d82cc3b6338a9425a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1271717bd6c4ad47ca5ab540:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd275452a7b4561d073c7d2f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85b9ff553f1ee51eee4b2ee0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:95e6d29eea890206b64a2b9e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a141648fd4d5e30c21cfe742:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6bc574010c1c62fdb5c06ab5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0bf41df877960a294bd2dc3b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:988af17e89f3dbd0c2a6d8c9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:996dc372ff31a1dbe5f80051:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:091c3f8e4c36f93849ac9785:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:079f9ee4be2ddb457026a050:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8bb5f7b4bbe51dc82621e3c4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:17b0baa120987bb826e6293d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0299d1ecf01b99211039b2c8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3a5c95f52cb7249931ae8054:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df8c537529a9f98e6d25c493:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d4eae14009ac454c6dd3050:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:062eac1ba14e29549c90605f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:782a9ce4f823249e70efd6c2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b1501a6246bc51a2aa4dd4d7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1d264e421eb3650ec6c88f72:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d292e83ed35deb88990c23a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b51485b0f6cecff843c3a3cf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a64958ccbaac405f2649562:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2bbca54f9fa3526037a45d5f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be9545bc7bad816745095a84:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d971b583c5f6549dda4a104b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0c431c169d5350813868fcc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c463c601d22ed6f862262e2:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d38c8882612470ce3b7dab3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1017887fed6b638c2029360a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:669b7a1ee069c220b8bdb64a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:73c1d9c12699beff2c57db5c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:604f087db4066c54036e358b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:518fe21c787ca08f888b605b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eaf3895d8703c0acec712ad9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba8ce6c41c9ecde628f8f085:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 572, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:060e5d024b462f4ae3b4f91f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 571, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf7e6ec22c3105e56dc1b43d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 571, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:97f6f5a77d290fa5591db5e1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 571, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c211443d389ac0b2eaf2e0f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 571, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d617576e2f876effec9bb0a2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca9adaf9abb1b02a390ba105:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b855b583ca6b1e9350d9aa73:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18833593ad322bc2dcbea547:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a969259678ed90e865c28fbe:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b4535265f0a2e35d06e4f570:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb48a578a464a1251337a945:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50948d9010b0cf9be4705043:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c6993e692c8ea421debb519:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c71a6020390573b60dc5648:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34e76f4638ada5d0d6ac7788:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cfb6b092cc0c84712b1aaf61:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3fd046e082b53d5371704ac:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:edd3c2b78bd679cff92b51ef:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e23870d7b4eda111ef3a4af:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:73b4b18f273a7d95d9e765ad:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c79eb4114c40e995f3c39e21:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65a2de6b84eca499474c8690:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:42b1e81510ad3aa3c14b7cff:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36c0a1b384208e83ada34096:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:977b3d8d36ad484117f35b86:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fc20dd3298a3c6bd8fcaf067:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3dc84da1587d06906545f61c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7cb0ac616386bc05c529e74d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8848bebe5d1b3ba8cfd9e0cc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6cc84949cb57ffcc2e00fdae:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8aea77496507afce37ecb5a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf5c447bb2d53e2bc6f8be9b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 573, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c405410739966da6329ca882:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 574, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4590b4784259a94ec3af0811:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 574, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a1837d609911a827f872b9e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 574, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea3dfdae3aef291e7700e753:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 574, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc3ae0450bc3fc8e68368bd1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b396bfa34a9d31d23deefa3d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7cbbd588f069c8555e06a379:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3237771955dd6b30d41e8c39:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6df4b63890e6a372f2df7e9d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c152aa8999fa73ff99ef885:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b025645c9360bf3a899387f5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f34cc937b036b479071a9d0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5acbbde21967398410a1ad3a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb9dc3ab77485e3a990e3d56:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bc767500ce31f92bebd87dc4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1616008ece3befbcf7ac3456:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:984309771153005679f20451:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba71bfed0a26b0af1bdd2695:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ccedacf25cac3995814020c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b126e98cf18d82316644705:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01718a17f99bed558550af79:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0fd7b0b208e08b9f3344eb52:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22a3b10970c4237e0c8ed51d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c5a6ad773b9dac3e0651374:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2d169c2d770eafd56fd4758:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 470, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3a59d88fc557dc149634d73:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 470, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdf26d139fbebaaa044c2fa5:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 470, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:055c56e08b4932985b34a473:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 470, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96571654b9ff81c52be288fa:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 471, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7196e90067afe6dd404174a6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 471, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:13b7758b6243cac5f4ce4f33:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 471, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29e879f91e5f944ae6fb86e9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 471, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:72d4a5c83ba0dc1ad7b77ffd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa1fa26103d5faf7c72e5608:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:074d4493dacd1dafc762a069:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e63620ad96cb32f54bf0a456:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 472, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b139a1805c6830a6e0346a13:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7028e58442b3c1ca5f7a2bab:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b48536a855b27c38c2b91f7b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1f1a3eb1f7ddbbdc865bcaf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bb7af29c3d9edb7c43eb5892:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4a4d7adebf309811c1e26a6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5d627257bd76f3bb531e065:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c8678893c0a76919aa83b1d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07501ccb8a07f49088214544:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e464a3a08ac314196aac74aa:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0027d23e3c9a1948f746974:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85b5063a4d5883992e6d8acb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f10fad0488934eb0b7414f32:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a56230a85cabeb26d046937d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7ea98cdff97172aa7909b3e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c6dbf140264ea67b36f0e22:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f253fa8955339167f5ef9e7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c63bbee2ca96de372c41de0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fbd1a1bae32dde189574541:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1be4d49642934fe08545764f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67f61a4add1339e5e069b739:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c346f8cacbacadb164856d6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bbcf2d453d8ff3bed96e9911:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e656e8fea427f33118c16387:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3f85c9e8efc59b4fcbbabcad:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ab01cf624903a190a6f9c45:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f5f538db7a194b4a2dd3e12:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6db135cb082032ee533b729d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b572035415c96f4ee0b9204d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7fd2cc20c2d4c0b6d47f2e1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0b994dc9230cf4cda2cb02e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50db9e62d29594e9b708142d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcf8541ce0550e598a5ba3dc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab8ae672e41046c1e9950315:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61a57e79294863efbaf08741:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e56b912c71d445003bfb4a5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:483889dfdd43b2b4b67d1f0b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:88c6ad09b0de5de5a96dc1a3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:618baa94527b6a8d7be08d3a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c84d1ea32238fa2cf7e9f3cf:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d44348b24ab9996c8fa75bd8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85f2721c164d94b1de3aae76:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c3a4ccc6de6097845cb1a04:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a5ebb92e1491f02dd8513c5:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b8dbbc2664243f529d57120:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5c96432e55ee99de9fb2b51:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:88f5725c37f73c24b68212cc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c26656968a0f61cf1c796605:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33faabbbfc6433ff4d689a22:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8bed7ccc4601dbf65880a3e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5cfb3ca0b69efdd6ffa3a4b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07307ec2d55120b44f4195f0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:287b4c2be4ce7f890ec460d4:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d19e27c3ab06efbaca54ed8b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c21c6ed160ba9b2d5d46b2d9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4987475bb3c5a76a9cbecaf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ff71f7ca6922896cf864dbe:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:545469addf9c3e20dc5ed800:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3da02d11ab3427b378923240:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12097124a965cffcf6e69a6e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b4c70a5822dcde19ef83d72:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0fcd4787a5d6c923bccfd47c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4dbf1a5fdadeb368ee2d6fdf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cffe2b5d1d8fae8977315f86:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:615fa4d6aab97aee94c4e8fc:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:147608a594478104c0d680e1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:45a81ded48065fad628676f3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8a1329ffe89c64471af3ff8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8fd926445bb1a824d991482c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ae74680c924ec9e11b897dec:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a538341eb215abf1380900a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc6390a98c2cf33ca8901588:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:341e910793242946a516b5d3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a451b75cbf44c73bf74336d4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab67bad44e9e5ecba285674c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1340a3c59769cdc31710736:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:600c4ef94cea891dc9f7fdb1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:959379badbee938823522893:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a958830f5b4e68527d62976:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f755b9e0473b052c0a3aa602:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b31a1d5a3ca6c179e0a8838:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dfe36dbe4acef0776455fc99:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40b79952844d2f3c619c1345:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fabe6d0c3e9a7212b5f81915:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f7598b99e691e09e3a1c0b0b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:94c7291587c7fdf8b9cd1f7f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d66f45b2ff8b346064a296e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e3d42be8eba9d63361c893f9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5ad7bbbe7642013833ec7e1:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5908e06d34bf003d7320046f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02a91adfcaedc94cd3fc583a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01ad27da1a36c9cfcfeee4ed:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 562, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75e80b41dac1e7e7e6338d60:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d89f7f3845a780149a83083:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6fd62684837c029a58c01961:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40070ccc7f7ea6b00c53822b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a118cd61fb57f5213613afac:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47f135bd19c3e53e991d8ae1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89a6d774cfc7356312c86566:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c689bc24e86a52674e7a022d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7fa60af1f417869daa465c4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07f166996e83a5c86e7d0708:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c721d20f39e8a865b3e38dbb:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a17f849dd131552ee8b1c0f7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c903d96368e8adb351b9fad:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f3b1fd80f7ecdb5a513b560:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7bc3aafe23a827ce7d229340:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c191e4c033dc516e2def8b7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f9b23c73cb7510a5a2adc3c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3fad815b9daffbe49eecf24f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:48a166fed1e0248e0dddab44:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:63f6cac5125a6e46b2e48347:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91e20cf6a48504b4ac254a01:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:519b1d4725a45bb41806f62e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c21cc954aaf7abf7c59f3b7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ea4788f85caaadb0bead233:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:493ec62203276581a79afaa6:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9837b7c810380c92f975e234:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3d525bb886f99fa5d17bfbac:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25d4490a9bce7ef88aa34685:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34db7cac091b269184b7468b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:24ec67ac5fd9e41a406e36bf:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7c85a52d4d1ee808facc872:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:32f8d457f532d2dfa2906b7b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 469, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:03c379704bf9c1ebaef44f82:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 468, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:781e96970e5b9263f95e9471:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 468, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5cf07cea038f1538c06dfbc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 468, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b23126dbe3242eefd1712c3d:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 468, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3212d633a8ddf2624acaa255:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3ccd8cc0a0446a25bd29e026:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c642da1c522c16fb79934626:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6780f90bd4ef0ab4c964c852:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7ce50cc4380b8a55d34f2ad:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4b222779e4ff9d85881cab3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2328e1873f3677593b45fe70:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3f8d269946cdd77250c0ca6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5bdf970cd0ba998b778b5677:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e33ab212ca290dda2901f5f1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a45b2db533539d94499ac60e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f50ed6b0f102353512c2c0f8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8e59a6e593d8418cb6710d4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22af5bbcd4d95a783b6dcc39:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:145fbd6ede95ce99e4bf7b3b:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e2e801e11135fd6f8613f59e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c3802c59affcaaade9387bd:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:945264ffb4ce0fa9ba9f62d9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:13ceb9c9e5fed7cd34292306:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c162b5b08fb13c438da4fcb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38df9ce8f0d3ae77cd5b3d34:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 557, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b3d3d83e371282b03a9d672e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 557, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3387f9c50b8eb2eb5bd70c80:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 557, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3b888f524c2b32f762e2701:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 557, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b20e8aff695f7d24e9be2b28:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9bcb34864cc835f2b791e60a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d028051d64f42217987da93c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87ea8523e8c98555a7446120:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:99abb8844b2fc5468bddb7d7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87c347f39283b0d43c77b36b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:485e1836d263485f014bf040:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f8688508369fe71529e15548:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:462244b37ace954a228fb0e6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75e7bc080a392a087b2314f0:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b43153d096085d3ad116033:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1aa4fde82e480d4ac83dc4c2:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0da2d71ea98b488d4825c359:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a054761dccb525a4701407c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:591f3eb85d17de8b1ac7f19d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ecaaa213a8ac9a46b6270ad3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cbfe4b454560682d4dac3674:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3d0301825674c044655bf054:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31d62825e4e3faabd7cc068d:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e981d07ed9b3bbbf0981a372:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b1ab97aefb5a673e3dc803e6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36772881471abb9fd040e64f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ad44eec89c2a9eb1f3f8eab4:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7fdee6f4441672fc1c465c9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:43364de707aa5df6deee5816:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1999401b40125ac07cad81b7:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca0dcd12281927da367a76f8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dda5d42b55829593d7fe9b0e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:727939aa5a329baf68d7a0e0:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f95b5688118480526695ea5e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc66325c50583ee4fdc23b20:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:59e119ff87abe2b42127574f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6b24c14151cba9f11c6e787:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:94a76bed42f3b0e80089d539:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f9fa00b0697ea252e9a99625:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd40fb140166100c77265ef3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 568, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:74a3b1206e2a1eab5a52bf1f:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cbaf99e91292389c5923b2fb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54a6e2e10c2f2cfc6aa6b776:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:afd58a1a0b150557c4f358b9:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 567, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61fe10a39e7ab28f99a50e4a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f52f2584a7c72fc9442beaf3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d914395c0130d80e25f5b1e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9adc6a9552d422d6d88b93e7:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22b5003ece8ebd2aa85dfb18:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:133375e6d8b9d78fde7fe0a9:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f55652b7259c7ee47df3119:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:163467daec81d95ca3d8af23:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ad54d23d660eb5e629f7f7c:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0af79b5ff2ab389db6a9b35a:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8388537ee2384b30c2c6feb1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdc7364b4884b8836a3504d6:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:833f6de9d608f30dab171f6e:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c7c2daa5474d5a1a8bf11fb:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:905af40eb3eb9c4955566266:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f18fb2abcbc23e0b64f52044:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61e1497ee9e585db875e4537:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:72db2db4d3ecbb56783e14cc:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7de5151d93cb1052dc0fa91c:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9aafc04b8b07637d4e6752b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c04098380a08cf3c5fd7dc4b:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:839128e9da684bba271045b8:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0cb40dc7be962f868cd0d15:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ff3ed1c28428f72ce3d54969:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30fd05c79c32add90cccda8a:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a80205837474bb0fd10f296:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9dec1ed34053d646cd07921:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa50726b01107030094d50a3:survive", + "split": "train", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:885ad0ef88ab51749e043fb1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f33b378decbed1eafe240310:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8efc83a4ba1f7cd1b0a3d137:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5324a7ccbb9088a70298bab3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:21bc249588c2781841c7149f:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5960e3b44b02676823bc81b3:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f886e82fd92bf0b8349c0131:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d595b4af79d59ee63a488656:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2908101dac2781f123f0ef68:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:628055a68b19320e37b5b9e1:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ef5a13f24bb691a08eb4894:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef5aa99df625925ff164df73:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2fb5cc2e2e671c63ecea93cd:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c75bd0440912fbec035278e:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c6963d8eeffcf39a12cdaf8:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:42c14eb9a1972ccc06eeb024:survive", + "split": "train", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f398ce00bd97fc5300437285:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c95a318ca2c09aab72f0b60c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7af16f45221c1b25d43fdc5:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b34dbf346315bea049b3fb3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96b9aa6cd719032a9c18257c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4ec420319578ef4ec44166b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50090cc679034c7daa5d5dff:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aeb3a33c603129f6ad48990c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4003b845d1dba349c5b5de1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3cf63a5a4dc396ebd6a1e86d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5dee3df69aa1c8dc59132a45:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:248b93ba85c9cbfb12c43dd7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33f56e9af08bb8c54919f196:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5de3d14d983a2614a765b5bd:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:46e83cc9f0b107b1d4b5dba2:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d8237b555d94e916dbdc00b0:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5663e783bb398ea72082df3:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c223dfab341eb69c79616729:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:46404f67802b9d293994222a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3fac0b725736aa037a284eab:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16e5d6ce016de554cdbd90a9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf1d97efbf2d8584141f71ed:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8951b6f604ec215762575314:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d052df36863e4df125f517e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ee2cd16ec08fe398bfa2a5f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c2385f928b4ed908721b40af:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4e3db884b0dfade44dd9330:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8409b7e8224ab1eb841b57b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e31cc834f61ecd9cb5c3e740:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f1dd8cb4fe6ab50fd4b3e11:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c02af18406fff8411250e95:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0979efce03ce9419d8994a1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed40ee9cca77a8be8b9fbbab:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5307030384d5fe47ab60330a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:707535ae8ed979160e6b04c3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb4c68dac851f72cb8fabc88:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0952ec78fc2d73f0df518fa:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9291fe8ba094d83b4573830e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c87e0c0516ac7f4fa71e77f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ff4cc1040f6229f89180a4a4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6b873ee7e0f5fdb1abf77ee:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41c2f6007e0e8a0bdb0e0f71:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3146e1cce7399b8e28959db7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58c6d4dd95cab700b72b193c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 201, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:576db2fc3e9ac93770de6483:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f1b104a5bf261b57ee60430:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d761557c925635868a3fc76b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3191d9616353be78e31fcb2e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bd78639e23fd2263538d1c0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b9d7b74fb4d78591c2e65eb:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b9b1854124a8236ba783096:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54d3d040008e83ee425b1f6f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5a17f6e42901b9f1d1d3c8f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:224571a455d751118a1592a0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f0acbc9db544f0da7b03b9e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81e6efa00f081652c8b096a4:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:96c11e9b0f4d5aef92569942:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:21935d79c6eb68227a4ad91a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f6937600ff421388af21a2e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:534fb0d488b7d85d398fbfa9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f0211ce19200b833f04fb7b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:310e0456c80d12247bf7d61f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4be135895b1292470b5613c7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:591c3cc1abd7e0baf57b6f56:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2dcbcfa3e1bf17697a7dc728:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5e23df8b5300f02cf83baa0b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df42f6342483435ef082cc08:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b581c4b5dd4a47ca358c58d0:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:00ad4502930c218eeaf66300:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:24598b2e1692167beb453b93:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6410bcc195012defc1780a3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f385b6f4f2ed1bb9687b6a26:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca558d1efedc7bc4c6d76a4c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5ac8020fe947017cb8dcc098:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:328a2a0aa9585e8b9d115453:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca5706b1e59c369c22972456:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:552728dca0751a5d929d866d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11f668887cb34681226e4eaf:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0276d049b418865881211dcd:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a5b02ca0a3827dd7e41cd7e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16ebb97c6418bf91877ddf82:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5a783eb6199ad75c4620232:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:429f555c250f7077556c6947:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bcd61b8ae841a07d4910f7f5:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e15fd669208dc30921c03369:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5455096fcc40f6e589bff2e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dd2ee3f2acfb76ab2546aa25:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ed526d0bc4f86b611cdb278:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f837f5e790c0ce258b9b477:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1bae6d5dfa00dea4a46ba23:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2538ecd39313177c5433cfad:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8423b15e5bd1044f29639afd:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e99b6575f88dec8d6c09f7d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ab549f45e7c611807b0c342:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31d8e8482febb32e62cf53a7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:49d20519fe9b492fd911b0eb:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6b15142fd829ac7e4583a9f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e768d2f79a8fe07650bdab4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d15a3dd3243ec9d39477ddf4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:de52fc939058cb7b8278b22b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6caf1df3ade6479c87147788:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:09541d962149d706515b7aa0:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16121dc0ed0ce041dc78f926:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:973d0d65ad51559ea03921b8:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87af4db526b7f881a4d5a315:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9d7a93f9a57dc7c27dd75e2e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:71c2e23c877efb17dc724b21:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01cbe0c184b8d989bdd66b8e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fd0c2983daf59eb4fb0300af:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fac77920146c5c8432abd7dd:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b54516c2ee133fb35b522bcb:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:675a9320f80d4b2f9fa53686:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:42a37ea561c0be6bc3533125:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89e0ba188ef8e6704a6d64dc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9cfdc0ced617220922eb71f6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39bd5569447e539c9e74aae2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a3df479a9e5186da2dddf934:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:19fa6556e3e5ffee61730514:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9d9a290f5daec2408097133:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8fe29088a567384e2e6bbea5:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5ea752773b4597b216c4798:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:952632a214ace986f3876d88:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:53d851e214605104686da4ae:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:436c2971fece8e8a1cd07b6c:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0820228a7229f63e299934c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:477d8ac01585dc486465ff0a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:306db963fb6d2627de3b7867:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7909d1e5faac480e7837c3f5:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fc59f5f020c78720bf8e749a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 260, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91efcc5af8dab70c42865e7b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 260, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c7c04444f50e40a1fe9bd0d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 260, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:551d132b7fe1957f1fafcd9c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 260, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f7e877788d4f0483fb831b4a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e1102b0e8c86b1205eb42ee1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:30c32511b5b558094e40f4be:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ede71c7b809f014437af8af9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 258, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b241cca34679e508f80d520:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9bda0666e6bfda7d18f928e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9b387744271fb301791a5ca:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:88d44fd1fcb501ac14a4ca3a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 259, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b78e91b08f7c33555a2be8e6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:337befead30095bfd00497f6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:331a356bf87970533c9d0725:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:497348247c579141e0dbb7cd:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5f4b5af955c3668ce46b278:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e9fb0462fdde24b92c4ae16:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:397953eb4614419d1829c2be:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e1354f3cfd523ca8a9ad26e6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ae476711536eeb7a36e22892:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dfcfc043a7fa144ae814b82c:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:679f14fc3de034a6be2d0bd0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:087bc9b0d1dcdf437027e20f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1e23bc5db13a97e887d68e0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:73e95c4d2684383be003f37e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87d722451f1b7193edb0faad:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:721ebd866cffbc7ef5060c48:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd75add80aa1f487583f4872:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c32bba374bdb7e14985c345b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a021259694317bdba5827742:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df6a5dd9c406bb8ebe3d304d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:555a65463db746a00caa9303:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a13771e185ce550ecceb9495:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ee0e3af1db8eab565e6c4cb:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:22fd55f789f1a68e66f445d9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:045fc472036e54cbb8252ae4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5ce09907c4bcffd15ffecda:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7f890a85ef1d61f9596fbf9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b44a56334a863cee1f199a4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:14974047485ab7c0046a98e3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a3b8ad471b72cb36bb601cc:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a89ad1dacc449a52d1ae6fba:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f89b9397e2f164a290b63b22:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 279, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c2006b576adbbdec1f51a99:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4265c05b34d7c35bd610aa2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c0690fd151f35e027f00ce7:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:43d345259f5826a942fdd763:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:305cbb6637f27234894cffad:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0aea7e42d966b3948d9e5bc7:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b67297d4c55629a3dfe7de3b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b055d9c44b56b5c3f4e70009:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:29f944e58ecb95c0cab6e537:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35893d776cf8a9befbf40035:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a9c32c7625e434330e609edc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6128e861a918690c6e5de3b8:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69f0be365acf463400f838f5:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62d7d2ddaa30a852b22711d8:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e5252df8fd788ab6798734a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc1cc99f8445a13669f31a12:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0a40e83beae3627e8be7e5e0:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:993de02d92df7259e27471dc:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11dd0449348645cb6381950f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2797e843c26bf42ebef56796:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b7c46a055482f10e1f7fc17:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a7264f7039f6beae0e9eb36b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:059b9e89e0f9a39e9ca07caa:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ccb8e1870088eeb3da0c8fde:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3cf32b9aab552f1682389315:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0324557b4afae51b6c3b6520:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:562d5fe09245bcb8558bc151:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d70aab02a83c9ca5067874b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 268, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e02e43a0c3c80a1601cfcac2:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e3f53db64493301d2eb03fc:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c68aab7c568da7cfe676b831:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0373c11329369971628bed04:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5931355aa7810ea88f612ff:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c8e6f7b20b98b0c3046f1e6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:44e50a716d69e952a59fe921:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:54cbbc305b7176514941a66b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b89665d9169a60779097a93:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ed998e08fde27f01c80102b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:685e864104193bc3279ffcf4:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dee9853401a8cbdb728be055:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f08c439be56316b9080dd5ba:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51280ec5eb510c0a97afadb7:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab2ce9be62657279cb62c903:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:511e9000cd444d45e17ebce2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b912e73239338bc21b705ab9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb133102174bcb7528b070a9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d56d6b7e687b0ccb678de193:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ff07fed19d0c89fc5f1249d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b5bab804434183269879087:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:84e155c3fec8100bdca752c1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d54d43a30b0adf2a51d7aec:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a10059fe1b4256a9ebee490d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5d2835e70cdf0207eb971a5:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 293, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:86124d379be3e4413450521d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 293, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:efe142c7a8eaec77652d4306:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 293, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6298ca633d87a99776608f29:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 293, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3cd6558cf84d7a5994a6622a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 292, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d76549b65f6aafcc5670f615:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 292, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80febaabbac7480b645587ff:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 292, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd157bd523cb73feed988157:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 292, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c667fbd78f30442caeb4e97:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:08e62a44f1cc4060da22ea9f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa3baf8ac48d9ffb0b7ff915:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:929e09f46361fa24a970c282:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 291, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:64f46f6ac1a317cf8acc09e4:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:09756350b0ec681b3a14174b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:88ed4ec705e6a442cc3f587d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50bf33865a0deab0365f116d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:37af3284cc151a723eabe51f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0f112aa4f4a68cd7eba0c37:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:57a0ec0dad63bcf13206c0fc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65c9e08f92a9ce202c0bee44:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51ce8ed8a12ba504bfe2ab70:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c915e3987dc51091da144fc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:660306ea3c618a2c8d964ba6:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51a70374061587647b530d28:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:323443dec2bfd8fc64c627c1:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea9b832e0022f5dfb2fb3c3e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bfe049b0a54b4bbe62048bf3:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07cc9ad2dbc6a87688551d5d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a47f17d3591dd007fe483f10:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23f5dfbe1d64b1cf8a522228:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:abe20fc8fce53fe9888b9c1d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5e60d2ddbb6468382c08734c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a80f96b87e8d723b92a77a10:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39919d49c9207c1e90f5a595:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:295a6fb411ad8c2a8ecf7d84:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23786cf4761e9b4b10e9cd4d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7daa4c4f395c335381c0d6d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2534b7357efb6be74fc48b7d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:798964b71191e9362595bc91:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d62d16366ccbd619783b8d59:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e727ac0c846d35db48a16604:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67cd960e6ca8f35adca973db:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a046ff2f9d09d7c2c33aaa20:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f8e25ac2ea9d85113b86073e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed206f27d5fd98e4879c9b15:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9aa1dbf07565bc86881ffc39:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d8c9dba76f0634f25015d92d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92d6713c5a1005b29650d9cc:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:155f6fd14641a2072ebe821b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9cf6b23995764fd03ee7bea:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dfcf1c40c9a2804319bb0b0a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28654d3ec6fb4502546b014a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:03f2ae04e713ab4a0d75ceab:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fabfd8e25fd21bed19be5b3e:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:63e1aae23481dce0a89b5b6c:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cad7109adbe1d0a97cb90b64:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:94ba63fc6f08206a978c264f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7299a5fb077300e87ab688d:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0d54ce421f99ef307f3f59ce:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:61e2ab96ceb285917e2797d0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75aad09329a738f0363ccf49:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:de65158bf80d5a38663dcec9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f768955b02feec7cfa83e201:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2aa90e678f7c83ea667bfd4f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7835a32e046860a028c75089:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:579cf9fe496f94d8a2ee08c6:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed75ec25f215b211bfbed7f1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:536db834b5b818bb16daa79c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 577, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a36bf8099a156adbddd2ad7a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7f3d2e4a645d95404112b30c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c0d19e415b0bcb3aca49fb9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89779b3d6e87b2f15c2de4c9:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d75e142ccb5692ea83e2aa4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8010f5ae742ae3fe2ea75565:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b05572e331ca11f13cb393f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bc5f6892031a4c76ee30c1b3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 578, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:53a373b84b65c14efd0a4836:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d789477e68b50e48e9a3f79:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:610f6229e6cda13d25a0b7ce:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e6a6bac193d16547aaf1849:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a41c002fcdd40680188ac946:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0241eb9082b557b013bab311:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bb89750127ac7be3b25d710a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a997f65f7e5cfe2a5a449f9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:84c07f9f155dcb320c251d6f:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1691659edde0695028e9dedb:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39f0c8cce80e33915eba1af9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbc4c721c2d6cdffbd79c471:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5e5fabcf3cae0eeeaddc500a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18cc15219fd4a1b47cda8924:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa51313c6a41f885e17b07c9:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6646e1db106cf350a236e6f0:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a6b65f9da451095a51adaf1:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:110e5571dfa4e32984d8bbed:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a13077063bb4b2e6453f258c:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:de9461d82cb7e70d2efc9a9e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d6315c1652269686febb597b:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65ef4deb55163817323437c5:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d07dcad3068d78d7e7caf1c4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1aad6d1c910112d41557409:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 565, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc8ca6089c697bc5a1492ca7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:545e4a8ad127d85f7be2b1a7:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9324a60519705f2e6ec6f50e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0084135d34eab3a7ddb40dc:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 566, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c16e7b978ef83f776bc0da4:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7384defad0b09055a7e5af93:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6e9827e301e4b6b4ff519bb:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ec7025b451dc2d96d979b36:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f57147da19294108965d0394:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38421e0749191c6415985d5a:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a2d8204e9a88523541f8227:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:682351d3410d2bea2fa6f1b2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2870f64eeb1b539fb2f77edb:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2a803e4987d80d76e1687195:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e56239e7fea917609bfd4e88:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f57263b7327c7058ad8556b:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eaf990b375eb6070abebbc27:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:76f41f8162d567eb515bef20:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8961abd3696162e3ad8fa607:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67ace2dd0074d6e92a9d8365:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e243aa382a403df328cae226:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2530b95b77a385e6843b03d3:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:084ed959138bf7612f20a416:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6387f001ebfa13c17ebd482a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ecb2114c65608c3b51e3a6d:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:786534c9af01c99a755be110:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2284f5093519d1c18fc2035e:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:882a6e804d22a30368d8c321:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0c7750a82db748fcbbd9219f:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:339a487d3651f5f4ea4d0b87:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47efb15edf8a829a6e32943a:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:657b760eef76afe98564a8af:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 414, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ad7bfa9a4d549142e5e05617:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1bd284a9cdb1f1050c9a503:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c72896206e66d0f8765d8e78:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e86b247e390217db561a58a8:survive", + "split": "dev", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8ea01478a316c176249ea340:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68b06f530c85fa183e07c6e2:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4af77e879dbdf5782dcde370:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba8ce1ae3aeebe4f43d17812:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93ba3b56fec917437ceebcb8:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f0dab6df61ca5ab3788ba126:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28c93c1f06a1d34b67e903a5:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81856ef4972db3d940df2c10:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5676985d18ae665f15ab884:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a36433584cc2c0e6be9185ca:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:709273537de2e62d3a52c427:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:654e7772810f847d66532c54:survive", + "split": "dev", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f24a3617fe02c91abbe5041f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2350d22c125a175fb10ca667:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc69b1c49aad3a88ccdcb851:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e233250fb2d6462001431874:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ba78f9f3c619329d795af5b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f145d6756d87026f8b6195c6:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58792b538006a2dac978a42d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cec5bed52d30e8e43a4548b4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b74b7c08ce380e74701f600d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:510ab59890d54ef432234f49:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b67ea8d1b41d8f2d5e8d54b4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8451388bc93fda2c517d8782:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d19a66aface410f373b064c0:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c3c4963d436bb3972057ac4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f3c306cf3ae7ebe371984e1b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2529c991ceddf93e1774ca1d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ddabb0077ef28acca27ca222:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a2d479f2af661bb1445796b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8cda16d57cdc7e27c9476e6b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf5b4d14e4dcf0727374be6a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b93fb286759b8e62292784a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:13353909aaacf548282639b0:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50d51c7ab2491e5f0bb3c92d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0366d43cd88b910555d9988d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee12a2ee10b48580d32943e3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7af336f36b376a6c77ca6842:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f05e1a2045f80e1d836c1581:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e71edd437a6a8a07d177b13:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f62ca6a2ed225e51db14225:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a5914416468517a0b0ff145:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d676b530ca5ed1ced22a0387:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf91d881fa5b9e0aaae9ea52:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62d6e02278abc36dc07ddab2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a017efb6234aa7b31af7d4b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:76ee37061a5405d319505fee:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:78fe39463e18193f6f722c20:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a5b7d707e00b38aea06af53f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed14cbcf4b9acf1423e88f23:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b8b1c5bef3c12a8de7c85e3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1aabc92cad14a9713c13a5eb:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf77734c437918e8f7e7731a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:74ae31ed6c8172c695484630:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1bd9842ee9d27d9fa47f8266:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e3b45aa18ab1b92efe48ee5:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fab8bb97fe11f1b62e8b9b27:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a098a7001506748903e34244:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3bd74fcf796dacaca3c1424b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b9afd90d8b5580a4cf108b7e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdd5f0a1e2ca707225df9d2c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:05235228d48d2c0f6e0a02f1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3827ae0d3110802458c785c2:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c1cc7ddb7b3ec2c242a9684:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e203b12dc1b7929794a55556:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:908f2ea15cb1a58118ee639e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e3f7b7be4d54ef85f042af0:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ce606021e7d267fb211dc32:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2fdeb459ae6aa66bcb33f4fd:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:71ecc43061baea0bb8d6f17b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:605a84bc61b79647bb6befd3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a3e0719675d7b131349bda7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06dfab5f3586102c06b11bd8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8e155e533cd3b40514de4cd2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:914f5793661bb16fba025e11:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c0614da34fe6bb649a6bccc:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1301ac3ce27d89bdfa895e5e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdacdc72817d55ab8b915247:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89cc9329d953fff42dc5511a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40c004c56e45ec837a9dde20:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8deb92b9ca849dc07976c082:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3cc0967c1105c10f9282b781:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1043ad4e1629659b9f22e73c:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:193473b8f7115a21adf5d14a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:86a1d90c6d74157ebf00f674:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ca5f56a805d7310b948bf64:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2fe55000dba0e9569483234:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06ffb824478d7940d6e87cee:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c74f3747bf72aa58fcc6362:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01278efe0ba68e26071eb7e3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b9c5c5aca5f7d162c33ba59:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e93cd98533b0ad7dc32081f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e986e02f7b5ac320d20e4db:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5bcabb6356565e568c2385f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e8fc97779ff6fc27ff2d45bc:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9e35df41f0ce4b91d15516e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a96974765448c82f9782697c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8766b052f48a9b2141f028ed:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df72184ed05171cb9855dbbf:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e110b6cd6d99c2b269d499b1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5129ad65df8c55a84df23027:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92b3976d280a542f9c9aef34:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ab53eb1f38044b52bae38c3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:422f9eaa4fe7c1f8a50ea2dd:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b7d11b16b688d51b539206f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef3060728bca1d4cf0f019a9:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8859b39384bb3674cf6be6bb:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c0555aaa0232ddd41d8e2cc:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3bf1133b5e36177f94ca1a0e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:707c76f4fc5aae7650e4e4fc:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:449fd5d4a4819694c23f312c:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fe60882766a2777ab2626e30:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb046d267656bd328719ec33:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:991ae9c245f3648bc6bb555e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fab016da4b8b42a7c1518243:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b065ecb382c44810956dd656:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e140c312293fc636872e53b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9157ea8f673273b0215af17:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:570df5aaa332d4a68c18bdf1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:488dcd8e14dbe73940196dc6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9f9f6be71a35fb5486446d3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:acdf1c88b4e2bef4ab7c7849:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:18d213a8bae7b3bccdf613e8:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51c409d6916c6d030ad1637e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5e4f8f1c65e2a9b22d376cb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:812d78c2eb35aa32c37a597e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d0b562dbcf35f1ea7a09744:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:27afb6de88ee9ec114ec0486:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb078bdfdc6ae4277678c522:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dcbb32bdadd494ebfadcfba8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d8952313ac103008ae74ad1:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e98e88936d9b8ac9043b78f8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25415ab0085df742e5084d07:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:87c192cfcdb1856fad3478c3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11c13a2daff536edc2fbbe36:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f10410c1e300c2ea71a05cd0:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 267, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d2803789969cf8d09304ba37:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f5547636f20990e8dca02ee:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d95b5182df9b6f9aa3613891:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:68cdf01d844d93287c153277:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:60814b3b9638b09ed683e08e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e8e8d0c6bc5136bc4ce5ede:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f2b89d0dbdda52ab6d6e50ff:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4473de6bfe4b7235222e5fb2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 266, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d56af286301a9a66175f01a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f413a7105a191f14f3fff3c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9edb6cfbdaa5a2c6de9aefd7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e59f601468f29e99099e93e8:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f4f926b73ab2335f7487d643:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50af27c07d00484fb5fdc2a4:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bbb7d0605872b41c51f51a7f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b93a7bf8525b63edf9ff956f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa9ab6774d73d5d7f080f96c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c8678b7d2c720fb65344a6e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d905e9d4b72f076d33c13d7e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab6e78fe497743f0c665271e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e299c8c41ebb74c6de1b84e7:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33cc9caa2c248519b00b234d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f1fd043e4bbabd2b700e5c8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3717b8258bc68faa8c07dde4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e10ac8eb3e3daed71a532fe:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a77ad1f41898319f0632a08d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e54fcfb4598d53919e450d2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0405e987ba41594ccabf7c73:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:775e102d75596b0afdbd90db:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fe6a66db0a05664fecbf76b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9358d6026a16d685e434f6c7:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2eb3e35689abf90e8bcbf390:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 281, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:64996692610d2a5b58ae7d13:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa647d7e2e1d8ea274a812cd:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ddd764f66f874821732ae62:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cc4d6270a2dbe0133403c5b6:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:628db00b8e82c227f96717e2:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb14ae09fc1f5737d251e397:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd5bca27b03de8f6c3efb042:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:512a26d67ae11e8d2c860f82:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81fd2f3cced5a25b0c96e652:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58534746f5ab71a986a774dc:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d64aac749500f9e584897d5e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e2863d514f130ed51f903779:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b72853b473ea0de9384cea26:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e0dbf6be9b6a683d91e6cf6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0cee46a52059780eb66ab99e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06790038d3806fcc21e42a9f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ff08fbabe8d5f4a1c81ec57:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:531c9d02df53393befafef5d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:169b51ac2e4533c2440c21e5:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc31237f03923622616a839b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b7283b6ec02cc568c4f5eb14:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fed65e8f556072b24338fd36:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66a6d74d15970f5e754937fb:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c83eb465fb0b7fd14c9277d5:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e70d82a90ad462e10db6ba3f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d897d668231ae80605f98de:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf23e82bf53e897e161d8837:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea4c7ab25c8f096b2dae54c0:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b278426dec26bc256dda4e75:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36eb18bf9f8cf102395d747e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c0a4cd9942e23c9e23c0a05:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8653fc3c4700ab9b3b7649ed:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c8d0b8e1bc9b24f82d07acf:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1b26ee01196e1ea7f07ea3c:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:04c467da98d7f56ad3c018dc:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a6a32cc5d567d9009b1f49bf:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 274, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58eb17212bae51b923a77c58:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b8594591afe004fef731dc8:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6c5b9ef3773c7e70f81d46e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eec5eb8d0f252bca0b15e523:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:83e304d67f5a1171ffc5a562:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2707bb2a79bc6e792efa79fa:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:522c14a39f5da560faf1a97b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34241971057b390340995342:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a41798edf526abeeb95b5363:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25f1f7e791ea0799b23ae83d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8420346b0e5a12ead19bca99:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:083954c9631613f6a35c706b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6af055c4464294cc4f119f3c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cb7fba40afa7b36a7cad11d3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5680dcedb9099c4244e49510:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2838d4c9fc818e92adacaaca:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3803ff8f185cb6c0b864b625:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e2263572d15ab9c42991d2e4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6964a194108140010a727229:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdd5dee44a8368a4fc90409d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5732eb3e29a444c6ee65874:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7faa9327efffa3dc14583067:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:443b0c18c2875f8a8e37a1f8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:77ede63aebd73d60183893a5:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:df8448608a68717291c0eb18:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5b49b6d4d2aa5d67b4cf944:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6634979ba0ea6ec8bbf26fba:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e564ae2439e35d264ee76006:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 278, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f4e7ed27454a23ea024d707:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:264cc44adff3e9f1fdf86480:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6bc0ef43c0ace87980e8503f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2307e0ee9f6e7a69fbb647ab:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 277, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7fe71e38c197f8e48fef8e48:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4cd42555038fba0422650601:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98c99c306daeb5bf3af9094c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee2489584e1bfc6af99707a7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac323cf2eb1ee685dfa522c7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c33dbf8ee24c8a70b5d50045:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd558108f5d2fbcab5bfa70f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7006d38bf86be5d8a59c9742:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:523595510e7296ebb20acf7b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf39e2bf39481de8f10256c6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7f413f388cfc14001fa5a824:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1d6b769c01f031cf30a8b17:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:728be71ed26f414e2a0ca7a5:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:45ae094c037d5fe0704671e9:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:891e1b52cd7ef73071aa5dcb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bd3ed5fb89e2c15c468a3fdb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1299983abee789571c9ac385:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f72357ec9461fa7521d7c9b4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b86a8e0188534972a6465778:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f685189d7a3f0217b9ed8fc3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40e67a322acc3fbbefcdafac:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8355d71eb417a7a94fd9f0c6:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25620227f87474990b837326:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ba24955414be2a069685756c:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a1a8ac0e48da71859fc97ff0:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9806f7229f492f708ead3da9:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c3bc051e04ce799f18b5654:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eedda693c65651a283918434:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ae764a9ca349c92d64a86586:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93f1ce93859b7cc9c5d8a81f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b975b51c04f7f4696fd7920:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c89642c2a50b9cf7f18e6002:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4148e4dc8b3a5e56083fbdc4:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbdaa8face17ddfe1a98b26f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2eb0e0b393805ef04c67907c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4fa8143ffc721f6280353a0e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fa26d6ae17994438f6c4ed9c:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:72a98a6bbb53a9f2fda94f60:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b67e800ba2a3e0a6eb459272:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35cfa778c07a952c2859155e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a43da79e448fa36e76ee3d1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:654ce403dd7e0f274a036ae1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a0020baa4fb513918b26339:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6254bf5d53b4be7132d2d910:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e39ea51203e2b96e7dd12341:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a55f8d812b545a871dac869a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d43103db8078f4d067ad3a1b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ce3cd2bb718f5276f6680a3a:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4f342a36cd26224a43d1bb5:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:099e0c7417dad965c8f917f2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2582f7bd8825d9c4af4801d2:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:03afd18817e9b25bfaa5b698:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9aa0c14bf7183cf4d1872346:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:317c4491ffba7f46d902c29d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d261c4848f9b77d95286b9a4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:edcac26c71b61ec69705eed3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ec66f54076ac77de1d51dd52:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c634002c241dc6fc71bde51a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9b11664acf4121adcdc67111:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a4f737bf5b5e0901b4538c6b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:82c075aad51ae5a7cb6af5a3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35c16e6d9badab247444d9a4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:407a51e349e7ec8b16b3f1be:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f845d257f2f3270388cbde8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d48825df0e1c4ac8bef6ec00:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:912ac89969d3782530f92967:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1047e0607d8e4830663d0505:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:14a876cbff1015cf972308d6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:64018ec56babe0609839e5a9:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7253fcf0764a46c7ff28248f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0103dd7c32ef82a1beee95e:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7a3acac21a9c4228a9ab0a1:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:836f0a0387bf6119f675f5a4:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:902cad43390a52291cde0b94:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6a1e6a0e0a2182d6f2c21c5:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d8b2d3b72fd44649717b06ad:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e049a91a38149f16d87f919:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab090279f03bfa45f1375e79:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ece7ae3f8666ed6a9184b183:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cbf0340aecd26990e0b8412:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aadab2875c22b9b2c62403e8:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 554, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a0aa98887d61798dd7a853fe:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 554, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6b30f34bddfe17b78849f021:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 554, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ecc2b752bb2631f83cb9f6cd:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 554, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12401d2736944514687c8ae7:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:48eaa4c95dd48f7beebf6fc3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cb7b1d761b288a522673f3e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4899b65e15f570ab23fac6a:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f846081c095ca7beba167a7:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7bd9a399059a2078b58da8c4:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:afafaa1fd9f895ced2c4c1ab:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa3b3cb3af6810aa3a31d11e:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 556, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be63bb38f41a4941d553d5a3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 555, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:717d8b1c969ea380f52afe36:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 555, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f73f895a308386d08adab8ca:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 555, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4768dad76749b886aaaa609:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 555, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fbcf752aab364fbf97acf0b7:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e6ea4f9d07c0de6c3691949f:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e51fdeb1a3c3c840b7dac1d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8bd8787839eff57d85091339:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d918756e233ac53f641ac325:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f19b01bc274a489cc430614:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f459fd3966f63f9bd9b08692:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6cbcd7834565e3ac20042685:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f178357f12791b2cdfdba3bb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3c52b7fb5554667a02023fce:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:63be0db1c2badf142dc7235d:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b4651e4262601b752ea2da4b:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:00e0fa31fd378d209ff756bb:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab1ce7b7f1b4874030601e80:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f1398f7918a12f8c420e511:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fdbe56852c5cddde7ede66b2:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6c933cb93ef12640810048f:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5cc5bf01719660c9e3942d1d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c66bdbc23fbc7bac9b650e74:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85dda89cf9227245bfe7946d:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ab0b66376a9bc06310642de:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1329356437645d7c33293744:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5bb04b7b365ede2a6b923669:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81200c3e10f262a29d7f8abf:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 570, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38b12c8ad0af9d6083610899:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a9fa002546c887a1e6052fde:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:838844fb9c237c11b0d67901:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f74612b21e17f245b6789018:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28a4b7f5c29ba1e8701d7a6b:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8bee79293ea193f85bd8f408:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c864c098b1f1dd7239a1bc3:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ca1aefea5abe484166e8d84:survive", + "split": "calibration", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 569, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7c79e74d67cd41d9608cd6b9:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a428124d2572b4ddc5f78b69:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6d82e1c9f5b568f5f96ba9d4:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f37d48c368380b33e429322:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9f2acfb91ca30ae535d98d8:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3db0e5eb4d0e8fad5ac9bd76:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5c5c9d6fbe0078b0b3c45b3:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d62dc2b322644af0c4b1433:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6bf98150b40fc673ae6f2e06:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:050eb5903da9dc589640c7ec:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41a75f7a84558f7ef2a91791:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:786721715802ac6831afefbf:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c588bf292216b33b8b19ec91:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:573b8a5483cc63b744071e42:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:77cd96f9eb526d87db0426bf:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e60a3473dbafa2db4894f1b6:survive", + "split": "calibration", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15a0a79ad1804a523d4fdcc3:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3aab322e68bc40267fc81e00:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0930f3bb683da1de730e12c6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9dec1b50aa908b6697c50c8b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:982a1f24d0672f692f3ab260:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9f263e16582be7736c59e60:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4eac891687aedc6e20a2fe4:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f5967e9bdb058ef58331e86:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c9aa936effa81886e4f8d2aa:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e53af6ec6f48e3b6ef308c30:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cf9e10a49a97e9e66cbedf88:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f6b1658a665077870054689a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6415756613b00d63f603dc5b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a1d2ee67666e9e76a433d98:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65d2a2a53f5eec3803ba833a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1aaccc663a04aa8228bacbc2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:355688f1d93ce79dbd3551e6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40e09f3769968df2baa702fd:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b8c53ddb9c12a4fe9446e92:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c87101a4925e298aa6ebbdd6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15690dcfcbeb905525378eea:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3269f1c6715be32f5918b43d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:81289dcf882179350140a4f4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd769d1328dfa811f037b9e0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5eee8c15777f36aa41d2432a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5ad4f998350d94b43e5cb66:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e6ad48460feae23637d2cda:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34c9876ee6537c7f90292664:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16347b895187d24626199382:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3f4f12649bfe560792742ba7:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a546f5f90b7b7ac01e1a4595:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ab7f7d5f4f3c302115c82baa:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8d8c7d1aab200648dee1ea3d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:310c7ba7ac4719062817d591:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:afd236258599e4a58de17166:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6d80797c7424475cbb46519c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d524de9e4d11c6f16c1a4037:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dc3844f44d7e3bcc429c3bc6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e99aea1e810e1e7b4c305923:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be6c3ff25d7cf8382b90f9dd:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:acd2a3c99c6af78c6e465be6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d7b6bb08c570fe4af3024659:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e11d7dcd3f35cc8663b8b81:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0bdadaea6ebd39a130e72e47:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:26a1681e7f60bc341323040a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:82203c0c43bf32f0d93c943f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c2e44d61fea9c77aa16d3be:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a9cee5dbf3a0ccad88aadfb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5d095b12fd7472d3bc8bb69:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31b9f493b908961e6811ab9d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:edd0e2ee17dd2da478eeecb4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b455ae9104c57566d6564252:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:640b059af09b84e9b7978903:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3247f454153089665ec9098b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e95e6a94af634c7109c3f246:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:62b69142175d94f3fd4e2d72:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f04a0e1c43053b82d55ecd0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb9458a3614f91257f274be5:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93fd401e3dc8e507c8a1464c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a4247841838fbbc31bfba19:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a98804b3a9a9bb3d16bbe974:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:279040d0da84d719e1008a7e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5013d3f84a766ff90e1d66d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:95a8f41f1a8abd2d3a9eb1a5:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28f66441157133b175210bdc:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2480546c3f0750ab8889063e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a6bdab83a4b7ecef3ef647c6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c72598e032df002a371c721:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ed030c57b48d07a26fe3fa94:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f49d1a1b8732118af3e1be90:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:41337ab27205460f2133b691:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:569e1b7cb9d952a5879b5568:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e40525b20d565be581c4ba0c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e5359bdc648fa50d4cac1891:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c29a5edeaf7b44106b0b134:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9c367aeeba3ed47df08cd733:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:866a5c381de4ac62118dff85:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23e4fdffb6c15673bcdfa30c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bf34ab76605a322f3022c5b2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e9dd8042ecdbb399e6c0049f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:465eb1946dce6dec3a2a5ac0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01d94f0b21a5dd2c818b3cd0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38f2113fc226de905ea3e002:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f67e5938b2605fee54199d36:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f9fac3a5c1238a6bab5e8372:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6978b8de7ca7784877ca9c1:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b4ee082c0adedfe3e83ab0a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aaec28883f5106f66e678b38:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4b666673a2b69725f3dfce6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1dea05de416e4b30a3d9b3bd:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e3b4aef7ed11d9fe77e2363:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:539335d4d742aa862a840934:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbb73448f12660f057c255be:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:69f213a32868f74881039073:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6f61de138da767b3a06938b2:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e34402918bf0bcc3917c54ba:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b8fc95d8e56831fb80726a1d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0809c5a5bf2a4a227982b51d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ff22df9dcf793090dd745cc6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:28b2c8e61c631a7343e23fe0:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a000537d614295fdf04d0998:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:131b6f3346e6a783bda21564:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bdf794b73f61d7e3a877b6da:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c0a5494be9df8886d8899ab5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:255ccea5a91c2b6559fce709:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8078bf5f14220044ed9caf41:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5798b04ac0704290fc794ed:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:49dd3e24e9d0ecd0bd6ca487:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:db25e3fbe8afafe5ede4a0f6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c28d78407447277ce418d6e2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b52a2679fd113557ab42323:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eb116adfca97f9e016d66bec:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7047e796c12bd31f40ca0d40:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f03be949c020e94742f9e77:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d25222ed9c8a3bd46b12cc20:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:463bd53f0721e822217ba46f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0ae468377d98732c59ce2f3:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:46479f8f0607160ac6dea63e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fb2996746d4a3a8542f6ff1d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:077f815d1122414cb8415a85:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7d573dd064a37b7a0c385b8d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c3412ff2ac67063d787d1c17:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:462618b61651570b2846c938:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f1183ebd02f47a05a9bc908b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e73459046bdc8819f7836a29:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ae6647d9b35a7839e4ef83e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67f1d172f24a2b42d4d9abb4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fff63af9fd2bd2bca03d9d72:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 290, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:637de0b4ce60d431b3fbe6b9:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4748d544ff79abfe5fab9847:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5a5948642f75f93c2f3cb77:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac5c668d5d11be78678d4a84:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:126562d789deb80356468d41:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:57a2008078f742ab6c8626e5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a4553b77d5b53b5790e3d20:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:11a4cd6425ef50288dd89dbb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c32ed72850524b6115259475:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:43b0d5e950a0b165bb220323:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f8fe2a27635c879fb8628e97:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ec1668afb0674d887d30b5cc:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbd521c43201e26268e3aa7a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93ce0e674e3a155e92ce6ff0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea556a3766a137afff9d5143:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d1c350beb0696f45d0b636d0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a561199f0e30a86517d4e77:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c7c26cd5a928391332a449a6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e112c4da03e89806f6b51f2:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d21883d41d3b3b5c5968af26:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b9715bdd36b25a38ad6486a8:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:820626b763f8825dd3068332:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:93d08bb7e0b93cb303105554:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e4a68ccc2d5e311494876e75:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 280, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d24a7c504ed2622641cdbead:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 283, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7cfdccb7443a93a634c433a1:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 283, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:817b57139b63e108f8790fdb:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 283, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d941b724bf9df118dd6a0aae:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 283, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0ee3c0f4347f3f58e953611e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3ffa2df372e8f6dfb0d8c38f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:471380d453ba14d8f9c03077:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5387a48bf735045df5884247:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2082df371acd1246e6974878:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e51957d703aa7a21a9126f7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0b27fc187d5b92f2f0baf02:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:388029ea4c6b9db0b05498d7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:596dc4bbff62dd637bf1abe5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0f95e6852ac5ed8170400816:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d53ad7b791e6754ee684dd46:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6c5abb646277c869e8f594f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a0e820d5880fd053a23b8742:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e24bd548aa26f5edb96ec44:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4020df98668c9b44c45f1f60:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a8f4fcc01b49eaa3bb6b30d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c8cdf2cb4c1de722f45f3936:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea686bc3ab05831b29e7008c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5c938b97813786d8c82237b3:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d649dddc9b41e2ca653328b1:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:07b6de7eb9bd7361c486b345:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:eef957bbb536e28fe4381bdb:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:19a7d6b2323b098d9e22032b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f2aafe130fc41373b6640f18:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7611787ac567633d94de55ed:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:648bbf706e496cd9f6908242:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0cd71dfc18adc305ffa5780d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:262929de5f678d99fd81523d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1cd8bbf14e4a59fe63113214:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4402a299cdebbaf033c9050f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f882ed980b69673ee6a7b58:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5d8322afa755ad70d59d60f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 284, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b959e9bf77a5adc5a30d2ce5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d189f2df0711b7fc2c771265:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c16014d87823dd357d1456d8:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:128f395ac37d7006e4e4c367:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c29119aae7c02211285f7082:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:697e27e971a1806c494ac95c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8047d9de8466175cbba13672:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5025ba4651167be2693e5f6a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f63abc827e81e246ac37e965:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a80ed0d47a0d3026ed1dfbc:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e7fe1b673f6177b0fa1dc76:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b1a6bb2f1ecd4b535f679a07:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6a50961b7fd5d31f3923d1b7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:441340c226bb1935b121e41c:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67f031bfe4e9e92d197132e1:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e0babcd3043ac68136a197e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5e495780c777233199b2eee:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8bd553e22df1683db52d5e78:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4892ab35e607d1b27a490c56:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a5f71abe56130d9b6d00a9e8:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:67be1599e86d96b0a79d4b3d:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c8656d105aa907f177cd2e4a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:501c446c78491b8f0a6b7edc:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:db5889f39922205969c462aa:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 287, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a67f9b5b6935e1e6b2573e9c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5b299de7d39edda771cb7973:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:342af4ee63b0cb125b271a4c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:625175dbc051d9cac87f6110:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 286, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6ac053036472f06ebe164a62:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51f4549cfb5680161d097026:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6231a699832f887534551fae:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9ec98f845e66e9a7221f2312:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 285, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f53aa74c3869112597b5fd30:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b13b8e3553d6936297b72580:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92cd14a08b1204b3c59807b6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5d250f37f1770efbf1655af6:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7bb180e591734d11d40b8605:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e7adbe36391fe00ae9a4164:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a73aabb2b12d1a995f9bcc66:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f5978de20565039401c0f9ac:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:814557a8578a6d65feb6e938:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35e4e09a850e616d48db8bfd:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33b52c598887d9d5905859f1:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:488ceed82540659725d8c36e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23c3f398425fa7ae16f59f98:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b7eb7993227f08865ceba7a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:851cc9d5fbda6c0c718c4fcb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:111c2d37ffdf960b57346949:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6c79b07c72f1575107a85cd9:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d41feeb008c6f9c5b2bacc21:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:896b4776fa05a557779d30c4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6ac74dc2de34b129cb9155a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:86965a2f8f90ff782961f060:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cef8be7cfcedc847fc2cfbf5:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b5469358631eb0c0fba97e23:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31afd4f6e7c73256fd123f75:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:50e9b66e55411fec32bf17d7:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e8b47dd39a6ad545fd84c1a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e915c93c80ef2f02d4d2c4b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8ff178390db9980396b19358:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 544, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1b6008f12b6d461b6f5ad121:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:55057bfba949c6ff0663fd0a:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fcf6ad681a5145718215a021:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:428aee950ad2683c813213a0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 543, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0b91f11fcd38154517858dc2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8259f446982e956a0ca11695:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d36a15fe27e140b6370c6f7d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:637cc81e55d6b29b3859c430:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a8baf32af0f2efafb0884ef1:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be4b7e8f2f86f141c2d8d3d4:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:51275459581e4b1ef3b00b7e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa2a4a1442ae92254da80c38:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dbc1dd0f7c711abf7adab9f5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1f8bb96e5dbcca559f3dec2b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4abdb09fd01605d136e6fb41:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:806d0448919413ca550c11d0:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8be2af3bc78bacaa2c393ca1:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2ed67e4159be3dfc4fbe7482:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:35fd0ba5d43edd45fd72ccd9:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9bd3410efc8104bc569c0c73:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f7fc6e2d78bdcaa8ebb70d1f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 588, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3ae94e3a3615b606c6b0faca:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 588, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:253b1f5b5850bacf600b161c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 588, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5ce3b5ce3aa46d05142aff02:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 588, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1ce51cdfd50612eeba18aa9b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 587, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0138be8aae7f889727aa1d1e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 587, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4b7e759881886577dad273fd:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 587, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ef6acc71d3fbb16066661281:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 587, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:27d8757c9da7f3430e5e0ced:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:194742a1a19aa36e85a58c7b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:31f3f29c21bbee6875a62524:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d7621d7b8893eedbe393db0e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4a3eb5c21fc375b3e82b339c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d10cb67b66d9c0fd81ac7e58:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1848c23cd4d35898d022bdc2:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a30ca432414e10d4fa53896f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 589, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6c30c88fd3483ab14dbaa1e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:16b40e61859c5318b4586afa:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8a0c1d6a2009f8c9b6913994:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a0bab24a3ae166837ddf068:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:40ab5b0e677bf9770e81090a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:78f00f7c7878440a2410f788:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f15638d409cc44fb52c8b7e4:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1131e5a097ac7e10af06232:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c2b532fd0a1c316f6a7827b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:def4366ded6a6df896671e94:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c25d1c92537c287c8d1f9ae:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5bb8c0d305fcc9190b60bd0d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3529da347672ddc6efcd7cbf:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3eeb11ec8c5b578dd385f823:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:55f8f4450093cca5395337b4:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4845c32b3bc94002c70ccfb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c29d1885fb3e711832008e6c:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:501e2691a84e6ddeeae2f9eb:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f293099110937a1af731b7c4:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:92488c2f6185ebce0ad9f995:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a07bbdfe3071266e2b3b2ca:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6595a8344e600006d2123f9b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2c4a4e147f7b1588ae0bad92:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4979e947e876782bb7b71363:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 563, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd3f7c9444d9e11b215839ce:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:866d5c488eed87d63b94c01b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b60f0028ea5cc952f66ae471:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7279558be989c68b7c5da9e7:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 564, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ec625bc4ecb9c56c9fa0cc5:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e4a7f385594c24fb238ae4b:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b6f2292977793d8d3b0bc2a0:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e331f261aeb0da55b58b86a9:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3e4af581b0e4f9fad15de3b2:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:102959b674a96c013936409a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:01f5cf08902258a28cdb1212:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6cf9cc6e93265277a8e3ed0d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f221f191c4d26720283f4d1f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6e11a42fce820771511feaac:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15450d6b94feffbbd82fe3f3:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d9e0925630cc8108b1dbf766:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6b98f2b0a9e5e50d70c9e675:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:241aa15d6210643b37819f49:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c8b4b84f80a8b426b676149:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:dfec15d15538d2133f64505b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:332c305d9f9064cca7c91b1f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c9d3d9f73d49180fd4e70c5:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8353849f306c99b801e76cbb:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0e5c16ccc37f603aaee2e0e8:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c5af1d3651ed77b0e4f433bb:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7b1cfff32e388417d93f789f:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f3fc69214f96664fd12c2681:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f10ee0a7b7901b7b0cffb768:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b1460dbef236498d111bc421:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1776662fd2b3d4233b7b1f84:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bbd1fa1192f750484e786f94:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8ecd72c6f944eace26d5a99e:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5a5dbb5ddf4c974b338731c6:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2afcaed2770a429bdfae7681:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4f0fd919b00922f0fb8004de:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91493ce1dfd1c9efd5084510:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c214eb63a496e85f4e8f0331:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e9860c60f475cb0c1bbc1ca:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b31bc0c29ed3b7f06b10dbda:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c73bfb4d052e52de134e6f20:survive", + "split": "test", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c31a0ea252ae23f905fd135b:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:305a069109d09eb613b4548f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e45ce175a6959ad79175b27e:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:656489c3fe1025652a510de3:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:91f5ffb49f77fc333df264ed:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:66e0e51ffb6f3943c3015dd7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c12252b469789af7f5190a79:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:563ecaa55e9433ce2d40fb0d:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c6ef55093396c3904f96ed65:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:924e4b96ca5999023e35d70a:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3b8e9ced855b770c0618a521:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:bfd687a18b43c9f071ab97c9:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:0902d059933889080b0053c7:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ee9468b38861c78cbeae36de:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1c77559f45064cd0b43746d8:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:85244a117edda6d6143e169f:survive", + "split": "test", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:65530be55a82ba6c9e4e0535:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2e11bf957181a9d4a1bdca5c:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6bb216eb099bbe2cb415eb1b:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c2a1c77b0784c41b332a688:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:aa2b06bff614037acff1d512:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7e9a16be72ced2fe22e4fcd4:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7df43449888199a23a8d880e:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2d7067b9c047e1fc87bb947d:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8f52a5c304a45097f71bf97c:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:63d57a7aeb929060cf51188e:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d0ed1675348397a71ec61d8b:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9e4f1cfcee5f63659228e6f6:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 800, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b53b47523823292cca2a20f0:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 801, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3bf115a8cf096d0947013ea4:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 801, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b0199d15491378b758402f5b:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 801, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:572f0ba1d3ddaa3fa4ac6f91:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 801, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c1e0301f1ecf0f1847123e18:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:040fc74e25d32f5a1706d9a4:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:12fef32144a2172110026e00:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:249c8645ae2858840c0d1c78:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39b9d4042a63f6ee6f7c2466:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ac51c42652a70e790e7111e4:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:23dd18b5e052cc51e1ba4052:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:47f13308a31f26e9f85206a8:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:39cfd2cf1e0faef8f4c0c05f:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:10589703550d91e77b38af7c:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5ff0bfd990d8f473981610b5:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e3d2edc1d02debf8b325bca7:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:90ead2aa54f0e923806f9b7a:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6266d716e135c92d2241af78:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:04100e1e91887aecab329285:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:640310b1795eba44df0405f7:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:823ed2dfef942950a0fe052d:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:89e8e93ebc09ab77a2d12ee1:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a2f03a7e02d4c7a0b37a5c12:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8df3da6709753e818e475063:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ec0ca15717c6beba10df0969:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:34e19699b6fbeb555cc4c3ec:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2498f981bea7f1ea2f94f562:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1181ce3fb218aa0ae26cd2ba:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:80ada521c29562b2d79f27b5:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d4d3051156c025b3e9c13127:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b17e8036bb65af3443e673ff:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:858226223a62dcae09bbb797:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1133, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8be9a77dcb5da3744baf8878:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:75d8ccffbb0e46bdf88c8778:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d5eee2c4c2557caa27de5ade:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:25092695911e018291bee049:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1134, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:249580a78256b5fd65542bbb:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:297c6e026ab9cf238b3c1d10:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b91dd50087e2e74714d9cf1a:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e56f2ee1a2eadd1013f0e3e8:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:942cd8d852efe24d6c5a4dd5:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:334d03453308ec3c45e12e23:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a267b801453dda4069a69b25:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:98025ad456cd47b9148284a1:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:90f596b7d7aa2159085c2b63:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:244f73488ae8b18c97afc783:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:15eb4c3530366489aea10f56:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9f0ea261f90dcddea760524a:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:db92f0e465c6399c5c4abec6:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:af3cc83d4f3621aacab12778:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e2293e4fd434a600da4d0bc2:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:cd59a2e1e889f95b39067f66:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c4997c1961582643cdb3747f:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:197e69acca81ee3796ea18df:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2b58c019eddd3d2c33f22959:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ef468fb72ac4cb64a81acb9:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:78eabe08f32e049a8cb4f8b4:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:48823570051dca86c519667a:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:27243d43f64937e1eb9b625a:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a9c17041a8da3f8c51d28132:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1123, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2f47dcdcc829ef63b4c6c5d0:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5f4459f112d4feb1760b4176:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:de191bbf708498615876e957:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4d76351a595e7c5c928de2ae:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06c1c1ac6665eca225c2382b:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fe46a01198229bfa376dd10e:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4e064899b758fedf86fa2005:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4318d2e14a9b1ce3c16c2894:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1124, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:f066a2f4cbc8579afd9f1982:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8c33d8efefc0fc8118834420:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:3bb9967a166f17cbcad10f3d:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9627aaf17a3806f0aedc0837:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7a252aaae386b04fc387afc9:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:38edf84bee3e6ad9c57f203e:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4c962093142dea21d0ec3dfd:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:4ca03f0754d5ce00a62bf0ca:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:196f3a6bd50e835d24bda265:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:2eef5b738aa0b07684bc1dd0:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:06f6dcb92ca8f002d5047baf:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:a477e556b4abb5cd224ffc4a:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b4683cbe2ce2d042e8341c67:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:05da9714c27d666caaa2a782:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:58ef310a92cb4b8b9b6a34ac:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:539c270926beb6b3742b164d:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:fab0fb117c6ad4e0061ee617:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:6df3c133fec6415c08fe51dc:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:753c524e8accedbf5235e6aa:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9a255272ba51fa1edc1aa73e:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:7ddd679c6df6a0939e230453:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8b23c7e2e9a40cbe74d0d0aa:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ca673d0199005992cdffce6c:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:031d1db0d38c30f30506a4bd:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:45b3799fabd153c39f7b6531:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8f743a364012858f43563b93:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:9229baa933c2fe26956e93c7:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:be37856b3136494b863c28fc:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1076, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5b1cfec010f249f02c487e0f:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:02a0a62903ef89c49f64ee06:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:e7dd94f0d108bbb8627c3b1f:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:8467efa8c94001e51b138cba:survive", + "split": "ood", + "family_id": "scaled_maze_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 1077, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:045010ec88340767d1b89b2b:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:13c6045871929ef1628066ba:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:36d14866217fafab8eae7780:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:1e7ed7df6d8739bb28b8e150:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:b49b03914fdf279abf49926f:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:141a5bed53a942c0406932ec:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ea6882511a6c6fdde770fc68:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c50529b6eec075502f964315:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:44bc47aee0dd3387580fbbe6:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:ebe5cf8672d9330665e9e010:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:538db869e56e44d7178b3be8:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:c841d71a8f7ea07ed3b92208:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:5db15071608cdf33b0e85961:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:d69cd166660da942fcd5aed6:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:97127e2f57c841d95a72d242:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "event_v4:33f2a2f998988627ddfa3f98:survive", + "split": "ood", + "family_id": "snake_noisy_actuator", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "programmatic_conditional_distribution", + "gold_label_kind": "observed_outcome", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + } +] diff --git a/variants/events_paired_seed17/tokenizer/chat_template.jinja b/variants/events_paired_seed17/tokenizer/chat_template.jinja new file mode 100644 index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4 --- /dev/null +++ b/variants/events_paired_seed17/tokenizer/chat_template.jinja @@ -0,0 +1,89 @@ +{%- if tools %} + {{- '<|im_start|>system\n' }} + {%- if messages[0].role == 'system' %} + {{- messages[0].content + '\n\n' }} + {%- endif %} + {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within XML tags:\n" }} + {%- for tool in tools %} + {{- "\n" }} + {{- tool | tojson }} + {%- endfor %} + {{- "\n\n\nFor each function call, return a json object with function name and arguments within XML tags:\n\n{\"name\": , \"arguments\": }\n<|im_end|>\n" }} +{%- else %} + {%- if messages[0].role == 'system' %} + {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }} + {%- endif %} +{%- endif %} +{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %} +{%- for message in messages[::-1] %} + {%- set index = (messages|length - 1) - loop.index0 %} + {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('') and message.content.endswith('')) %} + {%- set ns.multi_step_tool = false %} + {%- set ns.last_query_index = index %} + {%- endif %} +{%- endfor %} +{%- for message in messages %} + {%- if message.content is string %} + {%- set content = message.content %} + {%- else %} + {%- set content = '' %} + {%- endif %} + {%- if (message.role == "user") or (message.role == "system" and not loop.first) %} + {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }} + {%- elif message.role == "assistant" %} + {%- set reasoning_content = '' %} + {%- if message.reasoning_content is string %} + {%- set reasoning_content = message.reasoning_content %} + {%- else %} + {%- if '' in content %} + {%- set reasoning_content = content.split('')[0].rstrip('\n').split('')[-1].lstrip('\n') %} + {%- set content = content.split('')[-1].lstrip('\n') %} + {%- endif %} + {%- endif %} + {%- if loop.index0 > ns.last_query_index %} + {%- if loop.last or (not loop.last and reasoning_content) %} + {{- '<|im_start|>' + message.role + '\n\n' + reasoning_content.strip('\n') + '\n\n\n' + content.lstrip('\n') }} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- if message.tool_calls %} + {%- for tool_call in message.tool_calls %} + {%- if (loop.first and content) or (not loop.first) %} + {{- '\n' }} + {%- endif %} + {%- if tool_call.function %} + {%- set tool_call = tool_call.function %} + {%- endif %} + {{- '\n{"name": "' }} + {{- tool_call.name }} + {{- '", "arguments": ' }} + {%- if tool_call.arguments is string %} + {{- tool_call.arguments }} + {%- else %} + {{- tool_call.arguments | tojson }} + {%- endif %} + {{- '}\n' }} + {%- endfor %} + {%- endif %} + {{- '<|im_end|>\n' }} + {%- elif message.role == "tool" %} + {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %} + {{- '<|im_start|>user' }} + {%- endif %} + {{- '\n\n' }} + {{- content }} + {{- '\n' }} + {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %} + {{- '<|im_end|>\n' }} + {%- endif %} + {%- endif %} +{%- endfor %} +{%- if add_generation_prompt %} + {{- '<|im_start|>assistant\n' }} + {%- if enable_thinking is defined and enable_thinking is false %} + {{- '\n\n\n\n' }} + {%- endif %} +{%- endif %} \ No newline at end of file diff --git a/variants/events_paired_seed17/tokenizer/tokenizer.json b/variants/events_paired_seed17/tokenizer/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2 --- /dev/null +++ b/variants/events_paired_seed17/tokenizer/tokenizer.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506 +size 11422650 diff --git a/variants/events_paired_seed17/tokenizer/tokenizer_config.json b/variants/events_paired_seed17/tokenizer/tokenizer_config.json new file mode 100644 index 0000000000000000000000000000000000000000..9a87e3822403fb40a9524bb75edb66c6f0585750 --- /dev/null +++ b/variants/events_paired_seed17/tokenizer/tokenizer_config.json @@ -0,0 +1,30 @@ +{ + "add_prefix_space": false, + "backend": "tokenizers", + "bos_token": null, + "clean_up_tokenization_spaces": false, + "eos_token": "<|im_end|>", + "errors": "replace", + "extra_special_tokens": [ + "<|im_start|>", + "<|im_end|>", + "<|object_ref_start|>", + "<|object_ref_end|>", + "<|box_start|>", + "<|box_end|>", + "<|quad_start|>", + "<|quad_end|>", + "<|vision_start|>", + "<|vision_end|>", + "<|vision_pad|>", + "<|image_pad|>", + "<|video_pad|>" + ], + "is_local": true, + "local_files_only": true, + "model_max_length": 131072, + "pad_token": "<|endoftext|>", + "split_special_tokens": false, + "tokenizer_class": "Qwen2Tokenizer", + "unk_token": null +} diff --git a/variants/events_paired_seed17/train_log.json b/variants/events_paired_seed17/train_log.json new file mode 100644 index 0000000000000000000000000000000000000000..f8822a2e07b6a044eff467598411c72b1144fc61 --- /dev/null +++ b/variants/events_paired_seed17/train_log.json @@ -0,0 +1,964 @@ +[ + { + "step": 1, + "phase": "full", + "loss": -0.0748950969427824, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 1.1360169062390924, + "batch_question_ids_sha256": "9815e3084b0060ca1bf3f53e09caa2e8ba2ecfbfa18cd5a90c1b6ca1dc29fbcb" + }, + { + "step": 2, + "phase": "full", + "loss": 0.22397566214203835, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 1.9364449437707663, + "batch_question_ids_sha256": "0766e00bccd0b9d63535226a2a7454d88d591d6a1adb32befcd64d40174bd87c" + }, + { + "step": 3, + "phase": "full", + "loss": -0.009121095761656761, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 2.6859495593234897, + "batch_question_ids_sha256": "3eb8fb253ed8d59b98fd13fb12ae3aa08473ff98bf07f26c7537d8d4c6999318" + }, + { + "step": 4, + "phase": "full", + "loss": -0.04071702528744936, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 3.4203991554677486, + "batch_question_ids_sha256": "b72b9555459b9660b8cfea9c061b0d69e21b213181cf19973fa9312c1d740e4c" + }, + { + "step": 5, + "phase": "full", + "loss": -0.030815408565104008, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 4.176274714060128, + "batch_question_ids_sha256": "851370879eecc2bb2b872bf9f26a90b0c422dbb910faa7ab99625462dd70126c" + }, + { + "step": 6, + "phase": "full", + "loss": -0.003552184090949595, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 4.941503707319498, + "batch_question_ids_sha256": "a44ac9fef40754316a48915525e55ee3c8c5f613d2a3a206a70c53f6a360d6a1" + }, + { + "step": 7, + "phase": "full", + "loss": 0.005697006359696388, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 5.700752309523523, + "batch_question_ids_sha256": "dd15d5df027d50d8bc83cc139ce58f36adf2613954b88024320afffaf4f275ce" + }, + { + "step": 8, + "phase": "full", + "loss": 0.057425894298830826, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 6.455997716635466, + "batch_question_ids_sha256": "b16e21ce00030d172555d2ed47f3935798c5293130655f2d4bb114fab8d4f82e" + }, + { + "step": 9, + "phase": "full", + "loss": -0.03418128867633641, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 7.2420299937948585, + "batch_question_ids_sha256": "db13a8f636762d4e5c56f60dcecdc4a04f7acccee36ed4846fe242c110813e33" + }, + { + "step": 10, + "phase": "full", + "loss": 0.04701452422887087, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 8.022835675626993, + "batch_question_ids_sha256": "d29a2ab7bae9372a6460e6d41d2320b0a13b55633fcc21e2a06f06a64a2b8f22" + }, + { + "step": 11, + "phase": "full", + "loss": -0.004810761660337448, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 8.767681133002043, + "batch_question_ids_sha256": "e435c49e7cce06b5c387742c1cf2eae1fb3f57349a03ff3229713bf8e9179b7f" + }, + { + "step": 12, + "phase": "full", + "loss": 0.13739927113056183, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 9.533075327984989, + "batch_question_ids_sha256": "52067a44724ec88a3f85cace2ddba01f511bb396cbf27877697e85efd4f0c68e" + }, + { + "step": 13, + "phase": "full", + "loss": -0.07629979046760127, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 10.30644958280027, + "batch_question_ids_sha256": "d01e765c3bf9b2b1edf10d82d3b784f51514a21682d040badd6d9c1c29c3069f" + }, + { + "step": 14, + "phase": "full", + "loss": 0.06013527140021324, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 11.06470196787268, + "batch_question_ids_sha256": "f7d76860a7e7abbd079278179b2a6054aa8b1bf90c6e38db91f414cf906eff6a" + }, + { + "step": 15, + "phase": "full", + "loss": 0.12951497547328472, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 11.837062750943005, + "batch_question_ids_sha256": "1d089d97e3e6836ec1771dba1776d38f6a8ca414f96cc3c25fd0053ed3e27a44" + }, + { + "step": 16, + "phase": "full", + "loss": -0.11920184083282948, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 12.592718491330743, + "batch_question_ids_sha256": "abd204307499e62673daa334cd3a4533448a0ef971b517e7acbf23783905ae91" + }, + { + "step": 17, + "phase": "full", + "loss": -0.003990098484791815, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 13.34634671639651, + "batch_question_ids_sha256": "6d4029fa8044653ff0ccbe80e4371cccd55b6fc7c6462fb81d77f565bde1e2f5" + }, + { + "step": 18, + "phase": "full", + "loss": 0.0632672868669033, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 14.103745453059673, + "batch_question_ids_sha256": "88ddecb0e43b996acf8e321896454ceb6e79095126d68380cfdc7c74ba66d04a" + }, + { + "step": 19, + "phase": "full", + "loss": 0.042018274776637554, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 14.861994631588459, + "batch_question_ids_sha256": "3ba3b72125f7617c5cb4dfc031843bc10773d8aa7747ef6eef76ec813abc7a11" + }, + { + "step": 20, + "phase": "full", + "loss": 0.05378047423437238, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 15.614400346763432, + "batch_question_ids_sha256": "a8fbfc400a3dd61e2468ddde908d8e6914e8742e66df75fa209575bbd49f674f" + }, + { + "step": 21, + "phase": "full", + "loss": 0.020703650545328856, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 16.42179041542113, + "batch_question_ids_sha256": "49a7b1f6bb6891cd06f7c495942bbdbd98cfbd7435f49d77922ab364b9c1c5f5" + }, + { + "step": 22, + "phase": "full", + "loss": -0.010295026702806354, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 17.1975815044716, + "batch_question_ids_sha256": "d93e27a20a5666b8fe2ea5489dac0a58f99a770bfb0d4ee6870d36621ed78dee" + }, + { + "step": 23, + "phase": "full", + "loss": -0.031696483260020614, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 18.08850848581642, + "batch_question_ids_sha256": "88845418c486e4c7d856761b8aca680a7ca52ec59c84219e708745af76a445c2" + }, + { + "step": 24, + "phase": "full", + "loss": 0.0040627864527778, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 18.842839125543833, + "batch_question_ids_sha256": "b7fcb8d686d0a87ea4fd6ded60d595c7ccc4b132fc739e047564d72f5944577f" + }, + { + "step": 25, + "phase": "full", + "loss": 0.004738317336887121, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 19.609677392989397, + "batch_question_ids_sha256": "b2e6af990ea722e6cb276f57a736c23766306ae29b211bf00447d868b434c962" + }, + { + "step": 26, + "phase": "full", + "loss": 0.06255067884922028, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 20.376745360903442, + "batch_question_ids_sha256": "67253956a0835f33db1fa53443fed54476a027338e0ffdd7726dcb48fbef1c9a" + }, + { + "step": 27, + "phase": "full", + "loss": 0.06562656164169312, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 21.134697519242764, + "batch_question_ids_sha256": "b37cd333ba07e45e13875704fb0cc42b2f44fc5317737590bb27d88a7ce61452" + }, + { + "step": 28, + "phase": "full", + "loss": 0.03742673061788082, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 21.901406745426357, + "batch_question_ids_sha256": "d2a5a95e218ea89d7469dfdd8ff320de68f88ff52171675c0ccf7c9c3d95a891" + }, + { + "step": 29, + "phase": "full", + "loss": 0.006606193259358406, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 22.663256482221186, + "batch_question_ids_sha256": "1c6aa40b463139c53b7a93df96c3a1db9c53fac3123be7e8dda4fc8f53bc7996" + }, + { + "step": 30, + "phase": "full", + "loss": 0.0024183355271816254, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 23.401013052091002, + "batch_question_ids_sha256": "be09ca5b88c61e0617249157b2d75545b12e403abaf2ff570be1e6d30856d742" + }, + { + "step": 31, + "phase": "full", + "loss": -0.010469562374055386, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 24.170286952517927, + "batch_question_ids_sha256": "08f4544d1dfa624e78df3609c0059204a16146f8421bf85c14d0b10b5926b448" + }, + { + "step": 32, + "phase": "full", + "loss": -0.0184428533539176, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 24.91613855585456, + "batch_question_ids_sha256": "f8a0c95d4ce5a21d3b7bdd381eb7c9d15a251da10b33f0a8521a3e81eeb1450e" + }, + { + "step": 33, + "phase": "full", + "loss": -0.026352540589869022, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 25.68266553618014, + "batch_question_ids_sha256": "4c396fcb15e0a31d52d00473dc2f3bee3767da23146fb5038b6e63d9ad5016c8" + }, + { + "step": 34, + "phase": "full", + "loss": 0.005794679290374916, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 26.436185279861093, + "batch_question_ids_sha256": "c59ffed718f7ded62b1028779c65e8759de67070b26019e8642f75dd2e43a512" + }, + { + "step": 35, + "phase": "full", + "loss": 0.12788009410724044, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 27.187837212346494, + "batch_question_ids_sha256": "0e54ce4db6221fac98c15ce77d50701153708881dda86adbcb2776637215c9bd" + }, + { + "step": 36, + "phase": "full", + "loss": 0.09676597267389297, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 27.944907465949655, + "batch_question_ids_sha256": "24c257b893015e139aaec823c9fa8ed630a7d2df239552f229ebb832a7746723" + }, + { + "step": 37, + "phase": "full", + "loss": 0.04663453111425042, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 28.70325310435146, + "batch_question_ids_sha256": "bc22d73d0ac85948fd9bbd0813992b093f9beb230652c6234447dbbbfb4add35" + }, + { + "step": 38, + "phase": "full", + "loss": -0.06176069751381874, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 29.47435860335827, + "batch_question_ids_sha256": "6e47f7951fb42ce413ed2f4ce54d8832b534a3e9d0f2e35829a1608286227ef4" + }, + { + "step": 39, + "phase": "full", + "loss": 0.0276058129966259, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 30.218379349447787, + "batch_question_ids_sha256": "b5bfa6d3b4f6544cb42bc8b82b7c88ad3430367ce2f47dd0c3182d7f55e26796" + }, + { + "step": 40, + "phase": "full", + "loss": -0.026078838855028152, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 30.989923455752432, + "batch_question_ids_sha256": "e356d8692763396b1f305f967bc42bbea222b3adcfd7355c8ffd489d7c4a8702" + }, + { + "step": 41, + "phase": "full", + "loss": 0.04557723645120859, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 31.751174583099782, + "batch_question_ids_sha256": "eddf9bafc965ad35c8a9189e6a7b19eb336e5f2736ab490e5d0f7547f885b883" + }, + { + "step": 42, + "phase": "full", + "loss": -0.04995053354650736, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 32.51822272129357, + "batch_question_ids_sha256": "93f507828c7fa8715902d87edef912b05fe2920c25f400745b5d4b1835ff1fcd" + }, + { + "step": 43, + "phase": "full", + "loss": 0.09411330241709948, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 33.269474141299725, + "batch_question_ids_sha256": "50f72e83de9b1b9daf154f013aa536b1454d8d8c13b503ee6efa9007c721faa6" + }, + { + "step": 44, + "phase": "full", + "loss": 0.05266818264499307, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 34.04855671431869, + "batch_question_ids_sha256": "71067cb1ca0d666b5a16e53227ae1840e468c0c39e879ef8004ac73520b396f9" + }, + { + "step": 45, + "phase": "full", + "loss": -0.12074423860758543, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 34.81311255134642, + "batch_question_ids_sha256": "b9270e713e4fca02ecc7205c7ac1e320d3c87286eb7237ed0b02a4686b626bc4" + }, + { + "step": 46, + "phase": "full", + "loss": 0.001318799564614892, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 35.60654359217733, + "batch_question_ids_sha256": "fbecc4d2e806b5b42f4f82a7e99465e1fceb5b7b61dfc3cd167654deeef897d0" + }, + { + "step": 47, + "phase": "full", + "loss": -0.028849713504314423, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 36.41548177506775, + "batch_question_ids_sha256": "9fb1e3f8e3eb222e0e377eeedad5b0df90d11389465803b67a9b5090d2e79ca0" + }, + { + "step": 48, + "phase": "full", + "loss": -0.015318945050239563, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 37.20380338933319, + "batch_question_ids_sha256": "8ec776449fb8174bd55b829feff0711ca000a9e09d903e741da6f91e2d21a136" + }, + { + "step": 49, + "phase": "full", + "loss": -0.04532638657838106, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 38.08001133892685, + "batch_question_ids_sha256": "a20669c6cee5e157b8b54729bb88921f4ff6df89b4dda467ec25abb10f0e78da" + }, + { + "step": 50, + "phase": "full", + "loss": -0.04311649454757571, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 38.827290212735534, + "batch_question_ids_sha256": "cab2a2f34731e8a959300614046a8b8fc4c926971dd5909a1a6e3de0ad3acb26", + "dev": { + "questions": 372, + "objective": "observed_outcome", + "eligible_questions": 372, + "target_ce": 0.46014975992706547, + "target_kl": 0.46014975992706547, + "target_tv": 0.307618331572389, + "by_target_kind": { + "observed_outcome": { + "n": 372, + "ce": 0.46014975992706547, + "kl": 0.46014975992706547, + "tv": 0.307618331572389 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 372, + "ce": 0.4655643873714251, + "kl": 0.1955617947638946, + "tv": 0.19696661073944272 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + { + "step": 51, + "phase": "full", + "loss": 0.04189913161098957, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 45.038797207176685, + "batch_question_ids_sha256": "46758e1c6996a3f1a932a11e8372539949f8ac8f4c0415b36f65c713e195096f" + }, + { + "step": 52, + "phase": "full", + "loss": -0.07959715370088816, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 45.79904679581523, + "batch_question_ids_sha256": "d87fcd81d136c657c13e49aa033db193bfbb121e4a721fa86c16aad4c0a7c39d" + }, + { + "step": 53, + "phase": "full", + "loss": -0.051664914935827255, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 46.56836490239948, + "batch_question_ids_sha256": "4fae217aef8823502570d85ef34334cdfe2994ba567eef62990a60ac5e3ede4a" + }, + { + "step": 54, + "phase": "full", + "loss": 0.08312265900895, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 47.331904822960496, + "batch_question_ids_sha256": "edbcfec53ad977392c69172be6901bdb4eee1725bf8fd83582f64ac6320e77cd" + }, + { + "step": 55, + "phase": "full", + "loss": 0.011041590012609959, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 48.086882687173784, + "batch_question_ids_sha256": "7af8e2cece43ae333d4687839255ee144e5968fc9aa063efa2ff6dd96c42d1b2" + }, + { + "step": 56, + "phase": "full", + "loss": 0.13069539703428745, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 48.844940822571516, + "batch_question_ids_sha256": "6d579036d49f401569b6e29499a444725fe7a64de0907bc07bd389fd352e615b" + }, + { + "step": 57, + "phase": "full", + "loss": -0.058048512786626816, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 49.59209310915321, + "batch_question_ids_sha256": "848946e51c61b43a984fc49047cb8c9f0fae9f28c6447d8b4ee2c74dafe198bc" + }, + { + "step": 58, + "phase": "full", + "loss": 0.11841028090566397, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 50.3503486905247, + "batch_question_ids_sha256": "7a207cb4e3ff6157802e3b560ec0b63a488a8ab0eabaa5bcdea45eeeb107d3da" + }, + { + "step": 59, + "phase": "full", + "loss": 0.07705274596810341, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 51.090083728544414, + "batch_question_ids_sha256": "6ed68210f517203266a035e4dc739a5e91ad1773fb37b2d954aabbc97ce44cc5" + }, + { + "step": 60, + "phase": "full", + "loss": -0.01653895527124405, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 51.86966119892895, + "batch_question_ids_sha256": "cbce464330f46f2f323539fb9d7962cede14e7074be588012e1dfa684881e6d4" + }, + { + "step": 61, + "phase": "full", + "loss": 0.09852208383381367, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 52.62419832125306, + "batch_question_ids_sha256": "7212e07d655e85ec9b791cd6cde65ec3e9ebfa5224d1854c76d9a8e98cc07be3" + }, + { + "step": 62, + "phase": "full", + "loss": -0.06630048155784607, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 53.38278101198375, + "batch_question_ids_sha256": "768e8960dc65d1f0cc4b13cc2c90dbfe8db2c685cceb3a8ed4f834e59a13a2f5" + }, + { + "step": 63, + "phase": "full", + "loss": 0.00335497222840786, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 54.137751570902765, + "batch_question_ids_sha256": "ce0e1bf7240822fc1f373178c20604e9dba0409aa63f2b51415a5f4db012ea78" + }, + { + "step": 64, + "phase": "full", + "loss": -0.022465532645583153, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 54.904716577380896, + "batch_question_ids_sha256": "6b8c351645894c06829c8dffb3045f3a3081c0fcb0b24c924b9576410c78d47c" + }, + { + "step": 65, + "phase": "full", + "loss": -0.06888989172875881, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 55.66192784719169, + "batch_question_ids_sha256": "06de39fb8b5de5ce866344dcb10447ed159378857789cc67daf23061f43e3e7f" + }, + { + "step": 66, + "phase": "full", + "loss": -0.02314391452819109, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 56.42401862516999, + "batch_question_ids_sha256": "1728e00f1c3d6bfc533f2891740148ba51b40dcb766709fed2a21437282f6ee1" + }, + { + "step": 67, + "phase": "full", + "loss": -0.08001549076288939, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 57.179139980114996, + "batch_question_ids_sha256": "26b40bee228ef411ff536ea5c58c9a2fe4790fcc250d1ec97719f5c2b6051983" + }, + { + "step": 68, + "phase": "full", + "loss": 0.04526185989379883, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 57.95245848502964, + "batch_question_ids_sha256": "ebcd0c069c1aad6db15af178bfb86b95232c842c724cae991202c127dddf4df4" + }, + { + "step": 69, + "phase": "full", + "loss": 0.08819740545004606, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 58.71377477888018, + "batch_question_ids_sha256": "021f795640c8ebe5b45f63378c3fb38ba2c849f5d70b747f653482f046de3fba" + }, + { + "step": 70, + "phase": "full", + "loss": 0.15833008475601673, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 59.480925949290395, + "batch_question_ids_sha256": "968367da65a754ce0d7eb72abb8cbb64b558c86aa9d8adda10e119723caa2bb9" + }, + { + "step": 71, + "phase": "full", + "loss": -0.04195588454604149, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 60.24882358778268, + "batch_question_ids_sha256": "5f2d81f647af85f4a48f30464c10a61c6a85b7d5d35f687d527ade870b0c1e7e" + }, + { + "step": 72, + "phase": "full", + "loss": 0.22126729413866997, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 61.01563328132033, + "batch_question_ids_sha256": "6780ef3f330f6cd45a377b3f3c501ecc3ef2ae65d08769f207f4a68c152e3dbe" + }, + { + "step": 73, + "phase": "full", + "loss": -0.02661465760320425, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 61.79356468375772, + "batch_question_ids_sha256": "d521c674f97a615228053a262b84518189f9580e2c73b32eee51bcef5ed76d29" + }, + { + "step": 74, + "phase": "full", + "loss": -0.1296182470396161, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 62.70843842625618, + "batch_question_ids_sha256": "6c26d6bf906bf556525d65a2709f2f59a7cee21e2f62c8bad8c7015fcde12f89" + }, + { + "step": 75, + "phase": "full", + "loss": -0.01979048363864422, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 63.46672537922859, + "batch_question_ids_sha256": "afabf643cddd0692428615c4aa81345acbbcdab1c52a5c9a48c815f9e2700755" + }, + { + "step": 76, + "phase": "full", + "loss": 0.0012223836965858936, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 64.2400893131271, + "batch_question_ids_sha256": "ba8ce0da2f777532f12642355dcce1c4d29f8e0865010f26b4bf03c56d7723d1" + }, + { + "step": 77, + "phase": "full", + "loss": 0.12047487869858742, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 65.01171296648681, + "batch_question_ids_sha256": "9e0a4687654abac2570a41f8674ab95fcade2ac8e5e238d4661499dff7d58834" + }, + { + "step": 78, + "phase": "full", + "loss": 0.01269853487610817, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 65.77457793802023, + "batch_question_ids_sha256": "e36ae8ba0c5eaffb1a534d69421fb9118e4f45453e90b83cbbf5f29ef284cafc" + }, + { + "step": 79, + "phase": "full", + "loss": 0.15956797171384096, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 66.52514539007097, + "batch_question_ids_sha256": "171d14c1b2acfe9895a9353fe3dc68dcde6a998182a601a2746bf8fb10de20e3" + }, + { + "step": 80, + "phase": "full", + "loss": 0.021434137597680092, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 67.30584577657282, + "batch_question_ids_sha256": "5091eab21841dd9384edd2c719867375591caf9d084ba03c065878eb451a5674" + }, + { + "step": 81, + "phase": "full", + "loss": 0.08233899343758821, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 68.06916774995625, + "batch_question_ids_sha256": "f94cae69521603c0b60eb1cc6bb5370d2ea3d089f5b6bfd9a260b41a766ff974" + }, + { + "step": 82, + "phase": "full", + "loss": 0.06948617822490633, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 68.82234204933047, + "batch_question_ids_sha256": "f804d3278fd4930c24d09af7cc2e1fc70170c35cac45cbed9779d59f2d765e08" + }, + { + "step": 83, + "phase": "full", + "loss": 0.005867251195013523, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 69.58695920649916, + "batch_question_ids_sha256": "19b60c4f93e1ba7368b1f066c22c4ba32b8b66981562b2cbe2485da7433c07aa" + }, + { + "step": 84, + "phase": "full", + "loss": 0.08299335604533553, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 70.33971774112433, + "batch_question_ids_sha256": "ebdf0b28c90afc00f6691d1a09af18f9f66ebaf71377d0ba210fc4892869014c" + }, + { + "step": 85, + "phase": "full", + "loss": 0.011595956981182098, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 71.10882101673633, + "batch_question_ids_sha256": "4f04bf635d321e106f9856ba99a47fd6c2c033ae0450aa32c6dfe595a74ae4ed" + }, + { + "step": 86, + "phase": "full", + "loss": -0.03602656163275242, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 71.87254969682544, + "batch_question_ids_sha256": "ba9c22de6939f9c021221fe5a43ff5f50b15a70b21ed4effe8d97caf79b049ca" + }, + { + "step": 87, + "phase": "full", + "loss": -0.013593184761703014, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 72.63154661748558, + "batch_question_ids_sha256": "c9f35e67c5374cdfd4ecee262a8723f0f688d0387f073072324c9d259d9f4323" + }, + { + "step": 88, + "phase": "full", + "loss": 0.10619061812758446, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 73.39264491479844, + "batch_question_ids_sha256": "64ab5437c436a5c934bb05696c9b73f14c9838d3b210eeeb1114438303808400" + }, + { + "step": 89, + "phase": "full", + "loss": -0.054122308269143105, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 74.15655770525336, + "batch_question_ids_sha256": "6ff06537c7c8217ebac7a64762ff5ee340f367a0f7652d25d5cd799c0d15cbac" + }, + { + "step": 90, + "phase": "full", + "loss": 0.011591964401304722, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 74.92870382964611, + "batch_question_ids_sha256": "8fd09d36031bf8ee878da3b38a0694c5e416636fcd74e603f03dd205a9a3ab33" + }, + { + "step": 91, + "phase": "full", + "loss": -0.060869204346090555, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 75.69129009637982, + "batch_question_ids_sha256": "322fb22dffd39fb986e5bfd20ac78a9963d2d5b01978598e3ea9ff285c9b09d3" + }, + { + "step": 92, + "phase": "full", + "loss": -0.02303236722946167, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 76.43906010035425, + "batch_question_ids_sha256": "d0417dd44fb2608169bd7f2ab17d57acd203be7f8242f3c2e89ad317009a2e56" + }, + { + "step": 93, + "phase": "full", + "loss": 0.017207864671945572, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 77.20492906402797, + "batch_question_ids_sha256": "57be67994833e0a92fac3ba6c7b6b1fe3b8383d53284136aad0b9a926f82aa1f" + }, + { + "step": 94, + "phase": "full", + "loss": -0.04622429329901934, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 77.99955605249852, + "batch_question_ids_sha256": "9c3cbfa9806ac2c618d0a1e3ca3ef41c91b3a86c8aa45a36071383124ee71262" + }, + { + "step": 95, + "phase": "full", + "loss": -0.1110177468508482, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 78.75541434716433, + "batch_question_ids_sha256": "100fb8785fb8c78e98988f555871d3d9a4ed2716c9bff93597e8a64de5e2db46" + }, + { + "step": 96, + "phase": "full", + "loss": -0.015255212783813477, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 79.51458817347884, + "batch_question_ids_sha256": "c781c81a212d6bc7153d4ecdfb49ff5f6570c3bde61d9f7babaa983e433d7348" + }, + { + "step": 97, + "phase": "full", + "loss": -0.06440366432070732, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 80.28054185304791, + "batch_question_ids_sha256": "8b19c5c7751e07c22469266bfa40bb1eaac4670fe203a8c3e8a289350956893c" + }, + { + "step": 98, + "phase": "full", + "loss": -0.042902698274701834, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 81.04737775586545, + "batch_question_ids_sha256": "043573f3a457c1f8cfbcefabfef11e430a1dd573e779dd7e19f7dc26a4a1ebad" + }, + { + "step": 99, + "phase": "full", + "loss": -0.03668666258454323, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 81.81181224621832, + "batch_question_ids_sha256": "933b002848b57b1ae58768d4d38da9823ca19226389dff7c41f6ba3e58ff14b6" + }, + { + "step": 100, + "phase": "full", + "loss": -0.017029337119311094, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 82.66445402894169, + "batch_question_ids_sha256": "2a2b48544fe4380e81979729c21f00410faff377b407c4afdee1bef7530a71af", + "dev": { + "questions": 372, + "objective": "observed_outcome", + "eligible_questions": 372, + "target_ce": 0.45392147839038, + "target_kl": 0.45392147839038, + "target_tv": 0.2908940469690758, + "by_target_kind": { + "observed_outcome": { + "n": 372, + "ce": 0.45392147839038, + "kl": 0.45392147839038, + "tv": 0.2908940469690758 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 372, + "ce": 0.45846966947460877, + "kl": 0.18846707686707742, + "tv": 0.18377175030687837 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + } +] diff --git a/variants/games_api_seed17/backbone_config/config.json b/variants/games_api_seed17/backbone_config/config.json new file mode 100644 index 0000000000000000000000000000000000000000..b815be77563f3eef88baf4b4c8ae59706e618ad2 --- /dev/null +++ b/variants/games_api_seed17/backbone_config/config.json @@ -0,0 +1,63 @@ +{ + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "bos_token_id": 151643, + "dtype": "float32", + "eos_token_id": 151645, + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 40960, + "max_window_layers": 28, + "model_type": "qwen3", + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "pad_token_id": null, + "rms_norm_eps": 1e-06, + "rope_parameters": { + "rope_theta": 1000000, + "rope_type": "default" + }, + "sliding_window": null, + "tie_word_embeddings": true, + "transformers_version": "5.17.0", + "use_cache": false, + "use_sliding_window": false, + "vocab_size": 151936 +} diff --git a/variants/games_api_seed17/best.safetensors b/variants/games_api_seed17/best.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e4bf7f6ff6083bcb4bfb2475e64320ee60c72994 --- /dev/null +++ b/variants/games_api_seed17/best.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45829fc896391847220b67dcf6fc0702427d2f184be384649bcf9343c1552477 +size 2385039280 diff --git a/variants/games_api_seed17/config.json b/variants/games_api_seed17/config.json new file mode 100644 index 0000000000000000000000000000000000000000..8084b0e808ce7676ea533a1f6dc918641329484a --- /dev/null +++ b/variants/games_api_seed17/config.json @@ -0,0 +1,74 @@ +{ + "input": "data/scaled_games_labeled_v4", + "output_dir": "runs/games_api_seed17", + "model": "Qwen/Qwen3-0.6B", + "revision": "main", + "init_checkpoint": "/home/rwang/openjev_codex_20260917/runs/v3_teacher_coords_multi_seed17", + "objective": "teacher", + "loss": "ce", + "reward_samples": 32, + "gradient_checkpointing": true, + "set_head": "attention", + "steps": 150, + "head_steps": 0, + "batch_questions": 16, + "microbatch_questions": 4, + "max_microbatch_tokens": 16384, + "eval_every": 50, + "max_length": 8192, + "seed": 17, + "backbone_lr": 2e-05, + "head_lr": 0.0002, + "head_warmup_lr": 0.001, + "precision": "bf16", + "disable_native_triton": true, + "validate_only": false, + "self_check": false, + "schema_version": "openjev-decision-pipeline-v1", + "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "initialization": "local DecisionModel warm start, fresh optimizer", + "data_sha256": { + "data/scaled_games_labeled_v4/train.jsonl": "e21d05f1c34253687f4e7b78974503e7c05a7e145e4cbfc306246938d9f6317c", + "data/scaled_games_labeled_v4/dev.jsonl": "fd7780fd6bf94263f5653cea8b3575925447981602dba61d0c7ff96b8433ea88", + "data/scaled_games_labeled_v4/calibration.jsonl": "9632b050ee1069c4393e275fc878c2681f9c36a501c4c6584df19b3fb88d76fd", + "data/scaled_games_labeled_v4/test.jsonl": "82b62b39f56bcd703e6e7c911aafb20203ff7b48ebf4f4f4c3444694e4823fa0", + "data/scaled_games_labeled_v4/ood.jsonl": "568c11c100126b3848fb29c9cd3d4b8dc22be42fe2df43a3bb3cf8e7b0128ae1" + }, + "deps": { + "torch": "2.14.0", + "transformers": "5.17.0", + "safetensors": "0.8.0" + }, + "gpu": "NVIDIA A100-SXM4-80GB", + "parameter_storage": "float32", + "forward_autocast": "bfloat16", + "parameter_count": 596250498, + "train_questions": 1520, + "all_train_questions": 1531, + "schema_counts": { + "records": 589, + "questions_by_split": { + "train": 1531, + "dev": 504, + "calibration": 500, + "test": 487, + "ood": 173 + }, + "eligible_by_split_objective": { + "calibration/gold_distribution": 500, + "calibration/teacher": 500, + "dev/gold_distribution": 504, + "dev/teacher": 501, + "ood/gold_distribution": 173, + "ood/teacher": 172, + "test/gold_distribution": 487, + "test/teacher": 487, + "train/gold_distribution": 1531, + "train/teacher": 1520 + } + }, + "loss_description": "per-question proper loss; complete candidate normalization", + "rlcd_candidate": null, + "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection", + "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing" +} diff --git a/variants/games_api_seed17/summary.json b/variants/games_api_seed17/summary.json new file mode 100644 index 0000000000000000000000000000000000000000..0a09c5e7fa8f0fa7c5c99092469e7855d8d281de --- /dev/null +++ b/variants/games_api_seed17/summary.json @@ -0,0 +1,136 @@ +{ + "best_step": 150, + "best_dev_target_ce": 0.7130707709730839, + "selected_on": "dev target CE", + "objective": "teacher", + "metrics_by_split": { + "dev": { + "questions": 504, + "objective": "teacher", + "eligible_questions": 501, + "target_ce": 0.7130707709730839, + "target_kl": 0.10754575952110457, + "target_tv": 0.1622813057690467, + "by_target_kind": { + "teacher_rounded_proxy": { + "n": 501, + "ce": 0.7130707709730839, + "kl": 0.10754575952110457, + "tv": 0.1622813057690467 + } + }, + "independent_targets": { + "teacher": { + "n": 501, + "ce": 0.7130707709730839, + "kl": 0.10754575952110457, + "tv": 0.1622813057690467 + }, + "gold_distribution": { + "n": 504, + "ce": 0.701963093111963, + "kl": 0.6602369393657108, + "tv": 0.4357824048845307 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "calibration": { + "questions": 500, + "objective": "teacher", + "eligible_questions": 500, + "target_ce": 0.7210497784051056, + "target_kl": 0.11017916320286496, + "target_tv": 0.1606939957216772, + "by_target_kind": { + "teacher_rounded_proxy": { + "n": 500, + "ce": 0.7210497784051056, + "kl": 0.11017916320286496, + "tv": 0.1606939957216772 + } + }, + "independent_targets": { + "teacher": { + "n": 500, + "ce": 0.7210497784051056, + "kl": 0.11017916320286496, + "tv": 0.1606939957216772 + }, + "gold_distribution": { + "n": 500, + "ce": 0.690937608859615, + "kl": 0.6533763270403429, + "tv": 0.43010127156540956 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "test": { + "questions": 487, + "objective": "teacher", + "eligible_questions": 487, + "target_ce": 0.7178738686092758, + "target_kl": 0.11519962202078171, + "target_tv": 0.16197697068307357, + "by_target_kind": { + "teacher_rounded_proxy": { + "n": 487, + "ce": 0.7178738686092758, + "kl": 0.11519962202078171, + "tv": 0.16197697068307357 + } + }, + "independent_targets": { + "teacher": { + "n": 487, + "ce": 0.7178738686092758, + "kl": 0.11519962202078171, + "tv": 0.16197697068307357 + }, + "gold_distribution": { + "n": 487, + "ce": 0.6580319982288982, + "kl": 0.6173470163730734, + "tv": 0.4194799845862883 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "ood": { + "questions": 173, + "objective": "teacher", + "eligible_questions": 172, + "target_ce": 0.7490552275394107, + "target_kl": 0.10398421448117022, + "target_tv": 0.14334074401012564, + "by_target_kind": { + "teacher_rounded_proxy": { + "n": 172, + "ce": 0.7490552275394107, + "kl": 0.10398421448117022, + "tv": 0.14334074401012564 + } + }, + "independent_targets": { + "teacher": { + "n": 172, + "ce": 0.7490552275394107, + "kl": 0.10398421448117022, + "tv": 0.14334074401012564 + }, + "gold_distribution": { + "n": 173, + "ce": 0.7033095694567321, + "kl": 0.660899529493104, + "tv": 0.43063586291702377 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + "temperature": 1.0, + "temperature_fitted": false, + "training_seconds": 182.26331992074847, + "max_gpu_allocated_gb": 12.375407104 +} diff --git a/variants/games_api_seed17/target_audit.json b/variants/games_api_seed17/target_audit.json new file mode 100644 index 0000000000000000000000000000000000000000..3ac720324e267934ab6a829bcdcfd1fe9448d83e --- /dev/null +++ b/variants/games_api_seed17/target_audit.json @@ -0,0 +1,65809 @@ +[ + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.72, + "2": 0.11, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.16 + } + }, + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.69, + "true": 0.31 + } + }, + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.71, + "south": 0.29 + } + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.56, + "2": 0.26, + "3": 0.02, + "4": 0, + "5": 0, + "6": 0.16 + } + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.03, + "west": 0.91, + "north": 0.06 + } + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.55, + "2": 0.22, + "3": 0.04, + "4": 0.01, + "5": 0, + "6": 0.18 + } + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.78, + "true": 0.22 + } + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 151, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.19, + "south": 0.81 + } + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 132, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.33, + "2": 0.44, + "3": 0.04, + "4": 0, + "5": 0, + "6": 0.19 + } + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.85, + "true": 0.15 + } + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:45a206dbc4e330f637ece6ae:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.53, + "north": 0.32, + "south": 0.15 + } + }, + { + "id": "snake:45a206dbc4e330f637ece6ae:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:45a206dbc4e330f637ece6ae:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:45a206dbc4e330f637ece6ae:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:8d468090944737c21080f821:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.3, + "east": 0.64, + "south": 0.06 + } + }, + { + "id": "snake:8d468090944737c21080f821:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "snake:8d468090944737c21080f821:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "snake:8d468090944737c21080f821:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "snake:5d98b0d63d0ed94d96c88e17:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0, + "west": 0.54, + "south": 0.46 + } + }, + { + "id": "snake:5d98b0d63d0ed94d96c88e17:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.83, + "true": 0.17 + } + }, + { + "id": "snake:5d98b0d63d0ed94d96c88e17:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:5d98b0d63d0ed94d96c88e17:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.16, + "2": 0.55, + "3": 0.05, + "4": 0.01, + "5": 0.01, + "6": 0.22 + } + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6699999999999999, + "true": 0.33 + } + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.74, + "true": 0.26 + } + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.07, + "north": 0.93 + } + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.54, + "2": 0.14, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.31 + } + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.84, + "true": 0.16 + } + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.6, + "south": 0.01, + "north": 0.39 + } + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.53, + "2": 0.29, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.17 + } + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.78, + "true": 0.22 + } + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.14, + "north": 0.86 + } + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.66, + "2": 0.16, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.17 + } + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:da9d9f1b68c04ca029a5e799:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.11, + "east": 0.86, + "north": 0.03 + } + }, + { + "id": "snake:da9d9f1b68c04ca029a5e799:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:da9d9f1b68c04ca029a5e799:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:da9d9f1b68c04ca029a5e799:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:58d3c0dd72c2d2cd11756296:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.27, + "north": 0.45, + "south": 0.28 + } + }, + { + "id": "snake:58d3c0dd72c2d2cd11756296:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "snake:58d3c0dd72c2d2cd11756296:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "snake:58d3c0dd72c2d2cd11756296:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:07fcac0e43c5e1b61725bfd5:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.86, + "west": 0.13, + "east": 0.01 + } + }, + { + "id": "snake:07fcac0e43c5e1b61725bfd5:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.78, + "true": 0.22 + } + }, + { + "id": "snake:07fcac0e43c5e1b61725bfd5:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:07fcac0e43c5e1b61725bfd5:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "snake:16788c902da368b55922674d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.86, + "south": 0.12, + "north": 0.02 + } + }, + { + "id": "snake:16788c902da368b55922674d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "snake:16788c902da368b55922674d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "snake:16788c902da368b55922674d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.36, + "2": 0.38, + "3": 0.06, + "4": 0.01, + "5": 0.01, + "6": 0.18 + } + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.72, + "true": 0.28 + } + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8200000000000001, + "true": 0.18 + } + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 1, + "west": 0 + } + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.98, + "2": 0, + "3": 0, + "4": 0, + "5": 0, + "6": 0.02 + } + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.74, + "true": 0.26 + } + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.55, + "west": 0.03, + "south": 0.42 + } + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9999999999999999, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.09, + "2": 0.69, + "3": 0.05, + "4": 0.01, + "5": 0.01, + "6": 0.15 + } + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.65, + "west": 0.35 + } + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.23, + "2": 0.6, + "3": 0.03, + "4": 0.01, + "5": 0, + "6": 0.13 + } + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:1ba43d0958ba5992603084b9:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0, + "east": 0.75, + "north": 0.25 + } + }, + { + "id": "snake:1ba43d0958ba5992603084b9:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:1ba43d0958ba5992603084b9:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:1ba43d0958ba5992603084b9:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:bb33abaff323aa42615de283:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9999999999999999, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.41, + "north": 0.58, + "east": 0.01 + } + }, + { + "id": "snake:bb33abaff323aa42615de283:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:bb33abaff323aa42615de283:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "snake:bb33abaff323aa42615de283:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "snake:410f76de8b291ce18af3283b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 188, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.98, + "south": 0.01, + "north": 0.01 + } + }, + { + "id": "snake:410f76de8b291ce18af3283b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.75, + "true": 0.25 + } + }, + { + "id": "snake:410f76de8b291ce18af3283b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:410f76de8b291ce18af3283b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:3a37c45b8893e443bb48d780:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 188, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.45, + "north": 0.47, + "south": 0.08 + } + }, + { + "id": "snake:3a37c45b8893e443bb48d780:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6699999999999999, + "true": 0.33 + } + }, + { + "id": "snake:3a37c45b8893e443bb48d780:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:3a37c45b8893e443bb48d780:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 155, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.44, + "east": 0.56 + } + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.52, + "3": 0.34, + "4": 0.02, + "5": 0.01, + "6": 0.11 + } + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.9, + "true": 0.1 + } + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.84, + "true": 0.16 + } + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 155, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.1, + "south": 0.9 + } + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0.01, + "1": 0.91, + "2": 0.02, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.05 + } + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.86, + "true": 0.14 + } + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 155, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.38, + "east": 0.6, + "west": 0.02 + } + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.53, + "3": 0.31, + "4": 0.02, + "5": 0.01, + "6": 0.12 + } + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.88, + "true": 0.12 + } + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 155, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.11, + "east": 0.89 + } + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.59, + "3": 0.29, + "4": 0.01, + "5": 0, + "6": 0.1 + } + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8, + "true": 0.2 + } + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.91, + "true": 0.09 + } + }, + { + "id": "snake:4c9ad6e745c09cd665f24370:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.02, + "north": 0.7, + "east": 0.28 + } + }, + { + "id": "snake:4c9ad6e745c09cd665f24370:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:4c9ad6e745c09cd665f24370:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:4c9ad6e745c09cd665f24370:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:3e21643fee7f60ac15597733:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.8, + "west": 0, + "north": 0.2 + } + }, + { + "id": "snake:3e21643fee7f60ac15597733:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:3e21643fee7f60ac15597733:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:3e21643fee7f60ac15597733:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.9299999999999999, + "true": 0.07 + } + }, + { + "id": "snake:dc47b4d49347923b0a9e3af6:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.78, + "east": 0.01, + "north": 0.21 + } + }, + { + "id": "snake:dc47b4d49347923b0a9e3af6:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:dc47b4d49347923b0a9e3af6:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "snake:dc47b4d49347923b0a9e3af6:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "snake:adbaaf8a521d504bf37f192d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.17, + "south": 0.11, + "north": 0.72 + } + }, + { + "id": "snake:adbaaf8a521d504bf37f192d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.65, + "true": 0.35 + } + }, + { + "id": "snake:adbaaf8a521d504bf37f192d:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "snake:adbaaf8a521d504bf37f192d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 129, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.58, + "2": 0.26, + "3": 0.03, + "4": 0, + "5": 0, + "6": 0.13 + } + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6699999999999999, + "true": 0.33 + } + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.7, + "true": 0.3 + } + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 149, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 1, + "east": 0 + } + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 130, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "0": 0, + "1": 0.8, + "2": 0.1, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.08 + } + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.72, + "true": 0.28 + } + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 148, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.41, + "north": 0.04, + "south": 0.55 + } + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 129, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.19, + "2": 0.61, + "3": 0.05, + "4": 0.01, + "5": 0, + "6": 0.14 + } + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.72, + "true": 0.28 + } + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 149, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.8, + "east": 0.2 + } + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 130, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.43, + "2": 0.41, + "3": 0.03, + "4": 0, + "5": 0, + "6": 0.13 + } + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "snake:83f43ddfcbd1718e7a64288e:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.92, + "south": 0.07, + "north": 0.01 + } + }, + { + "id": "snake:83f43ddfcbd1718e7a64288e:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:83f43ddfcbd1718e7a64288e:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:83f43ddfcbd1718e7a64288e:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:5b6fc4a98ac6c78a118ac241:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.3, + "east": 0.07, + "south": 0.63 + } + }, + { + "id": "snake:5b6fc4a98ac6c78a118ac241:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:5b6fc4a98ac6c78a118ac241:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.64, + "true": 0.36 + } + }, + { + "id": "snake:5b6fc4a98ac6c78a118ac241:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "snake:caf8e8d695dc780f1b381231:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.76, + "north": 0.06, + "south": 0.18 + } + }, + { + "id": "snake:caf8e8d695dc780f1b381231:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.89, + "true": 0.11 + } + }, + { + "id": "snake:caf8e8d695dc780f1b381231:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:caf8e8d695dc780f1b381231:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "snake:881dea5763a938372f85c67c:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.59, + "south": 0.36, + "north": 0.05 + } + }, + { + "id": "snake:881dea5763a938372f85c67c:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.75, + "true": 0.25 + } + }, + { + "id": "snake:881dea5763a938372f85c67c:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "snake:881dea5763a938372f85c67c:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 130, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.61, + "2": 0.17, + "3": 0.02, + "4": 0.01, + "5": 0, + "6": 0.19 + } + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.78, + "true": 0.22 + } + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 148, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0, + "east": 1 + } + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 129, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0.01, + "1": 0.96, + "2": 0, + "3": 0, + "4": 0, + "5": 0, + "6": 0.03 + } + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.85, + "true": 0.15 + } + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.49, + "east": 0.5, + "west": 0.01 + } + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.38, + "2": 0.36, + "3": 0.03, + "4": 0, + "5": 0, + "6": 0.23 + } + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.71, + "west": 0.29 + } + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.58, + "2": 0.22, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.19 + } + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:e04aa1b9b164091e5b605d77:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.59, + "north": 0.2, + "south": 0.21 + } + }, + { + "id": "snake:e04aa1b9b164091e5b605d77:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:e04aa1b9b164091e5b605d77:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:e04aa1b9b164091e5b605d77:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:07fc2dce6ac4c6390adfb38b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.44, + "south": 0.55, + "east": 0.01 + } + }, + { + "id": "snake:07fc2dce6ac4c6390adfb38b:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:07fc2dce6ac4c6390adfb38b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:07fc2dce6ac4c6390adfb38b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "snake:cf4d4a371cc2b84d602a4bb9:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.79, + "north": 0.03, + "south": 0.18 + } + }, + { + "id": "snake:cf4d4a371cc2b84d602a4bb9:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "snake:cf4d4a371cc2b84d602a4bb9:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.75, + "true": 0.25 + } + }, + { + "id": "snake:cf4d4a371cc2b84d602a4bb9:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:ebfaca907fc41ff93ba56e2d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.1, + "north": 0.09, + "south": 0.8099999999999999 + } + }, + { + "id": "snake:ebfaca907fc41ff93ba56e2d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "snake:ebfaca907fc41ff93ba56e2d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.85, + "true": 0.15 + } + }, + { + "id": "snake:ebfaca907fc41ff93ba56e2d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 132, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.53, + "2": 0.09, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.37 + } + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.81, + "true": 0.19 + } + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0, + "west": 1 + } + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0.01, + "1": 0.88, + "2": 0.02, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.08 + } + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.26, + "west": 0.05, + "south": 0.69 + } + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.39, + "2": 0.45, + "3": 0.04, + "4": 0, + "5": 0, + "6": 0.12 + } + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.89, + "north": 0.11 + } + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.25, + "2": 0.64, + "3": 0.02, + "4": 0, + "5": 0, + "6": 0.09 + } + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.79, + "true": 0.21 + } + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:5d4c1f3be8e78fae1f7c577b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.92, + "west": 0, + "east": 0.08 + } + }, + { + "id": "snake:5d4c1f3be8e78fae1f7c577b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:5d4c1f3be8e78fae1f7c577b:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "snake:5d4c1f3be8e78fae1f7c577b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.9299999999999999, + "true": 0.07 + } + }, + { + "id": "snake:49cba9d9430dfafd8026f045:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.02, + "east": 0.51, + "north": 0.47 + } + }, + { + "id": "snake:49cba9d9430dfafd8026f045:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "snake:49cba9d9430dfafd8026f045:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:49cba9d9430dfafd8026f045:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "snake:89f4318abbbf35dae0a9d5fc:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.83, + "west": 0.15, + "east": 0.02 + } + }, + { + "id": "snake:89f4318abbbf35dae0a9d5fc:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "snake:89f4318abbbf35dae0a9d5fc:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:89f4318abbbf35dae0a9d5fc:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 144, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.37, + "2": 0.22, + "3": 0.08, + "4": 0.01, + "5": 0.01, + "6": 0.31 + } + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.9, + "true": 0.1 + } + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.85, + "true": 0.15 + } + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.77, + "true": 0.23 + } + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 158, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.95, + "west": 0.05 + } + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 145, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.65, + "2": 0.07, + "3": 0.02, + "4": 0, + "5": 0, + "6": 0.26 + } + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.81, + "true": 0.19 + } + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 158, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.64, + "north": 0.08, + "east": 0.02, + "south": 0.26 + } + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 145, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "0": 0, + "1": 0.22, + "2": 0.55, + "3": 0.08, + "4": 0.01, + "5": 0, + "6": 0.13 + } + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 158, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.7, + "east": 0.3 + } + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 145, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.44, + "3": 0.45999999999999996, + "4": 0.02, + "5": 0.01, + "6": 0.07 + } + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.87, + "true": 0.13 + } + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:d1757a09cb988028246a2278:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.02, + "north": 0.86, + "east": 0.12 + } + }, + { + "id": "snake:d1757a09cb988028246a2278:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:d1757a09cb988028246a2278:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:d1757a09cb988028246a2278:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:1f94c03e232d5cecf420b4b0:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0, + "south": 0.99, + "west": 0.01 + } + }, + { + "id": "snake:1f94c03e232d5cecf420b4b0:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "snake:1f94c03e232d5cecf420b4b0:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:1f94c03e232d5cecf420b4b0:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "snake:9bc6c48d7b6609943aa22be4:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 1, + "west": 0, + "east": 0 + } + }, + { + "id": "snake:9bc6c48d7b6609943aa22be4:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:9bc6c48d7b6609943aa22be4:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "snake:9bc6c48d7b6609943aa22be4:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.47, + "2": 0.26, + "3": 0.06, + "4": 0.01, + "5": 0.01, + "6": 0.19 + } + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.79, + "true": 0.21 + } + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 156, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.03, + "north": 0.97 + } + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 143, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.78, + "2": 0.02, + "3": 0, + "4": 0, + "5": 0, + "6": 0.2 + } + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.83, + "true": 0.17 + } + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 157, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.64, + "west": 0.34, + "east": 0.02 + } + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 144, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.2, + "2": 0.47000000000000003, + "3": 0.06, + "4": 0.01, + "5": 0.01, + "6": 0.25 + } + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 155, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.06, + "north": 0.94 + } + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.61, + "2": 0.16, + "3": 0.01, + "4": 0.01, + "5": 0, + "6": 0.21 + } + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:0994b5c61a691e7ea285be3f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.31, + "south": 0.66, + "north": 0.03 + } + }, + { + "id": "snake:0994b5c61a691e7ea285be3f:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:0994b5c61a691e7ea285be3f:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:0994b5c61a691e7ea285be3f:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:cfd599b961f4135d3c87387f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.01, + "west": 0.91, + "south": 0.08 + } + }, + { + "id": "snake:cfd599b961f4135d3c87387f:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "snake:cfd599b961f4135d3c87387f:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:cfd599b961f4135d3c87387f:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "snake:39b76b489b61357e661386a2:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.15, + "east": 0.5, + "north": 0.35 + } + }, + { + "id": "snake:39b76b489b61357e661386a2:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "snake:39b76b489b61357e661386a2:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:39b76b489b61357e661386a2:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:3643892fdb0609a4df6b640d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.76, + "west": 0.01, + "east": 0.23 + } + }, + { + "id": "snake:3643892fdb0609a4df6b640d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:3643892fdb0609a4df6b640d:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "snake:3643892fdb0609a4df6b640d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.06, + "2": 0.79, + "3": 0.05, + "4": 0, + "5": 0, + "6": 0.1 + } + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.72, + "true": 0.28 + } + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8200000000000001, + "true": 0.18 + } + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.01, + "west": 0, + "east": 0.99 + } + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.93, + "2": 0.03, + "3": 0, + "4": 0, + "5": 0, + "6": 0.04 + } + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.16, + "east": 0.84 + } + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "0": 0, + "1": 0.04, + "2": 0.8, + "3": 0.05, + "4": 0, + "5": 0, + "6": 0.1 + } + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.65, + "true": 0.35 + } + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 1, + "west": 0 + } + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.79, + "2": 0.14, + "3": 0.02, + "4": 0, + "5": 0, + "6": 0.05 + } + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.64, + "true": 0.36 + } + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "snake:f0189682c2eb6935a5d15fe8:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.29, + "west": 0.6900000000000001, + "east": 0.02 + } + }, + { + "id": "snake:f0189682c2eb6935a5d15fe8:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "snake:f0189682c2eb6935a5d15fe8:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:f0189682c2eb6935a5d15fe8:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "snake:981ca032260a3dffcec1bf99:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.94, + "south": 0.01, + "east": 0.05 + } + }, + { + "id": "snake:981ca032260a3dffcec1bf99:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "snake:981ca032260a3dffcec1bf99:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6599999999999999, + "true": 0.34 + } + }, + { + "id": "snake:981ca032260a3dffcec1bf99:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.18, + "2": 0.6, + "3": 0.06, + "4": 0.01, + "5": 0, + "6": 0.15 + } + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.78, + "true": 0.22 + } + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 1, + "north": 0 + } + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.89, + "2": 0.01, + "3": 0, + "4": 0, + "5": 0, + "6": 0.1 + } + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.83, + "true": 0.17 + } + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.26, + "south": 0.71, + "east": 0.03 + } + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.16, + "2": 0.63, + "3": 0.05, + "4": 0.01, + "5": 0, + "6": 0.15 + } + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.92, + "north": 0.08 + } + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "0": 0, + "1": 0.27, + "2": 0.56, + "3": 0.02, + "4": 0, + "5": 0, + "6": 0.14 + } + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.78, + "true": 0.22 + } + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:373cfe9faa2058b0490d322a:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.01, + "south": 0.11, + "east": 0.88 + } + }, + { + "id": "snake:373cfe9faa2058b0490d322a:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:373cfe9faa2058b0490d322a:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:373cfe9faa2058b0490d322a:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:16756b2a682710731d258d4b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 1, + "south": 0, + "north": 0 + } + }, + { + "id": "snake:16756b2a682710731d258d4b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:16756b2a682710731d258d4b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:16756b2a682710731d258d4b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "snake:696202fd3a2ce99a2445985e:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.5, + "south": 0.32, + "north": 0.18 + } + }, + { + "id": "snake:696202fd3a2ce99a2445985e:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:696202fd3a2ce99a2445985e:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:696202fd3a2ce99a2445985e:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "snake:4a9cd87f6087d543fb114185:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.66, + "east": 0.3, + "south": 0.04 + } + }, + { + "id": "snake:4a9cd87f6087d543fb114185:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "snake:4a9cd87f6087d543fb114185:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:4a9cd87f6087d543fb114185:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 156, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.69, + "south": 0.31 + } + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 143, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.56, + "2": 0.22, + "3": 0.03, + "4": 0.01, + "5": 0, + "6": 0.18 + } + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.91, + "true": 0.09 + } + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 157, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.02, + "west": 0.98 + } + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 144, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.84, + "2": 0.04, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.11 + } + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8, + "true": 0.2 + } + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 155, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.68, + "south": 0.05, + "east": 0.27 + } + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.29, + "2": 0.55, + "3": 0.06, + "4": 0, + "5": 0, + "6": 0.1 + } + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.91, + "true": 0.09 + } + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 157, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.02, + "south": 0.04, + "west": 0.58, + "north": 0.36 + } + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 144, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.05, + "2": 0.66, + "3": 0.12, + "4": 0.01, + "5": 0.01, + "6": 0.15 + } + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "snake:a1a646f11d996660cf23de69:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.01, + "east": 0.16, + "north": 0.83 + } + }, + { + "id": "snake:a1a646f11d996660cf23de69:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:a1a646f11d996660cf23de69:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:a1a646f11d996660cf23de69:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:54d4c05b4976562ce121bf21:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.32, + "west": 0.65, + "east": 0.03 + } + }, + { + "id": "snake:54d4c05b4976562ce121bf21:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:54d4c05b4976562ce121bf21:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "snake:54d4c05b4976562ce121bf21:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "snake:9d823e17bb4ea3f97c12577e:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.77, + "north": 0.22, + "west": 0.01 + } + }, + { + "id": "snake:9d823e17bb4ea3f97c12577e:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.69, + "true": 0.31 + } + }, + { + "id": "snake:9d823e17bb4ea3f97c12577e:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:9d823e17bb4ea3f97c12577e:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.92, + "true": 0.08 + } + }, + { + "id": "snake:56e8bd28a696b0a31a95926f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.76, + "south": 0.19, + "north": 0.05 + } + }, + { + "id": "snake:56e8bd28a696b0a31a95926f:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:56e8bd28a696b0a31a95926f:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:56e8bd28a696b0a31a95926f:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 225, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.27999999999999997, + "3": 0.39999999999999997, + "4": 0.1, + "5": 0.02, + "6": 0.19 + } + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 253, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 253, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 253, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 253, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.01, + "north": 0.99 + } + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 224, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.06, + "2": 0.67, + "3": 0.05, + "4": 0.02, + "5": 0.01, + "6": 0.19 + } + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 237, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.09, + "east": 0.68, + "south": 0.23 + } + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 224, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.47, + "2": 0.33, + "3": 0.04, + "4": 0.01, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.31, + "south": 0.69 + } + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9999999999999999, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.57, + "2": 0.21, + "3": 0.03, + "4": 0.01, + "5": 0, + "6": 0.18 + } + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.65, + "true": 0.35 + } + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "snake:1a3e61f06ce2e98ebb505ccb:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.58, + "south": 0, + "north": 0.42 + } + }, + { + "id": "snake:1a3e61f06ce2e98ebb505ccb:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:1a3e61f06ce2e98ebb505ccb:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "snake:1a3e61f06ce2e98ebb505ccb:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:e4df2f7dacb5671625749f3e:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.54, + "east": 0.02, + "south": 0.44 + } + }, + { + "id": "snake:e4df2f7dacb5671625749f3e:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "snake:e4df2f7dacb5671625749f3e:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:e4df2f7dacb5671625749f3e:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "snake:edf573117fe7de0a60f09ec4:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 195, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9999999999999999, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.29, + "west": 0.7, + "east": 0.01 + } + }, + { + "id": "snake:edf573117fe7de0a60f09ec4:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:edf573117fe7de0a60f09ec4:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "snake:edf573117fe7de0a60f09ec4:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "snake:f3ded4511a05512af9c015ed:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 186, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.68, + "north": 0.3, + "south": 0.02 + } + }, + { + "id": "snake:f3ded4511a05512af9c015ed:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "snake:f3ded4511a05512af9c015ed:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "snake:f3ded4511a05512af9c015ed:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.09, + "2": 0.54, + "3": 0.22, + "4": 0.03, + "5": 0, + "6": 0.12 + } + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6599999999999999, + "true": 0.34 + } + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.97, + "west": 0.03 + } + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 197, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.68, + "2": 0.12, + "3": 0.05, + "4": 0.01, + "5": 0, + "6": 0.14 + } + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 219, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.03, + "east": 0.68, + "north": 0.29 + } + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.21, + "3": 0.6, + "4": 0.08, + "5": 0.01, + "6": 0.1 + } + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.21, + "north": 0.79 + } + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 197, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.4, + "3": 0.44, + "4": 0.04, + "5": 0.01, + "6": 0.1 + } + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.65, + "true": 0.35 + } + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "snake:63eb6eaccebe64181ab30b1a:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0, + "south": 0, + "east": 1 + } + }, + { + "id": "snake:63eb6eaccebe64181ab30b1a:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:63eb6eaccebe64181ab30b1a:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "snake:63eb6eaccebe64181ab30b1a:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:19d04caecfe12a6a9c24e521:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.9, + "east": 0.08, + "west": 0.02 + } + }, + { + "id": "snake:19d04caecfe12a6a9c24e521:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:19d04caecfe12a6a9c24e521:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:19d04caecfe12a6a9c24e521:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:5eea186d00c75f1dfa595738:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.7, + "west": 0.27, + "north": 0.03 + } + }, + { + "id": "snake:5eea186d00c75f1dfa595738:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "snake:5eea186d00c75f1dfa595738:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "snake:5eea186d00c75f1dfa595738:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:067ceb1f04a6877293ed2723:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.32, + "east": 0.05, + "south": 0.63 + } + }, + { + "id": "snake:067ceb1f04a6877293ed2723:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:067ceb1f04a6877293ed2723:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:067ceb1f04a6877293ed2723:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.26, + "3": 0.53, + "4": 0.09, + "5": 0.01, + "6": 0.11 + } + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.63, + "true": 0.37 + } + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.06, + "south": 0.94 + } + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "0": 0, + "1": 0.22, + "2": 0.54, + "3": 0.08, + "4": 0.01, + "5": 0, + "6": 0.14 + } + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.84, + "true": 0.16 + } + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 227, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.94, + "south": 0.01, + "east": 0.05 + } + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.36, + "3": 0.44, + "4": 0.06, + "5": 0.01, + "6": 0.12 + } + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 224, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.64, + "north": 0.36 + } + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.4, + "3": 0.48, + "4": 0.04, + "5": 0, + "6": 0.08 + } + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.73, + "true": 0.27 + } + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "snake:d176ad19885d837f1a7982fa:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.19, + "south": 0.13, + "east": 0.68 + } + }, + { + "id": "snake:d176ad19885d837f1a7982fa:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:d176ad19885d837f1a7982fa:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:d176ad19885d837f1a7982fa:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:0efeba2cbe230918df7006fa:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.19, + "west": 0.05, + "north": 0.76 + } + }, + { + "id": "snake:0efeba2cbe230918df7006fa:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:0efeba2cbe230918df7006fa:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "snake:0efeba2cbe230918df7006fa:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:94c72585efdea8dab81e3ed6:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "west": 0.2, + "south": 0.68, + "east": 0.11 + } + }, + { + "id": "snake:94c72585efdea8dab81e3ed6:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:94c72585efdea8dab81e3ed6:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:94c72585efdea8dab81e3ed6:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:e1e15ce94bc3c638ab2c3eb0:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.64, + "east": 0.04, + "north": 0.32 + } + }, + { + "id": "snake:e1e15ce94bc3c638ab2c3eb0:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:e1e15ce94bc3c638ab2c3eb0:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "snake:e1e15ce94bc3c638ab2c3eb0:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 235, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.63, + "south": 0.37 + } + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.28, + "3": 0.62, + "4": 0.05, + "5": 0, + "6": 0.05 + } + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8, + "true": 0.2 + } + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 235, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.31, + "north": 0.64, + "west": 0.05 + } + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.34, + "2": 0.47, + "3": 0.13, + "4": 0.01, + "5": 0, + "6": 0.05 + } + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 236, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.08, + "east": 0.02, + "west": 0.9 + } + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.27, + "3": 0.62, + "4": 0.06, + "5": 0, + "6": 0.05 + } + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 235, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.09, + "south": 0.91 + } + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.23, + "3": 0.67, + "4": 0.05, + "5": 0, + "6": 0.05 + } + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:72dfb5d2d802b721e1e7ce85:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.1, + "north": 0.2, + "east": 0.7 + } + }, + { + "id": "snake:72dfb5d2d802b721e1e7ce85:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:72dfb5d2d802b721e1e7ce85:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:72dfb5d2d802b721e1e7ce85:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:5e8e60860af25ef0e831dd82:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.66, + "north": 0.31, + "east": 0.03 + } + }, + { + "id": "snake:5e8e60860af25ef0e831dd82:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:5e8e60860af25ef0e831dd82:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "snake:5e8e60860af25ef0e831dd82:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "snake:96fc807e107c37d4a70d690a:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.34, + "east": 0.04, + "west": 0.62 + } + }, + { + "id": "snake:96fc807e107c37d4a70d690a:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:96fc807e107c37d4a70d690a:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:96fc807e107c37d4a70d690a:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.49, + "2": 0.27, + "3": 0.09, + "4": 0.02, + "5": 0, + "6": 0.13 + } + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.84, + "south": 0.16 + } + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.38, + "3": 0.45, + "4": 0.06, + "5": 0.01, + "6": 0.09 + } + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.65, + "true": 0.35 + } + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 219, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.46, + "east": 0.04, + "south": 0.5 + } + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.45, + "2": 0.41, + "3": 0.04, + "4": 0.01, + "5": 0, + "6": 0.09 + } + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.8, + "west": 0.2 + } + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.51, + "2": 0.3, + "3": 0.05, + "4": 0.01, + "5": 0, + "6": 0.13 + } + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6699999999999999, + "true": 0.33 + } + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "snake:89f007044a81d170cbb9af1c:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.05, + "south": 0.04, + "east": 0.91 + } + }, + { + "id": "snake:89f007044a81d170cbb9af1c:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:89f007044a81d170cbb9af1c:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:89f007044a81d170cbb9af1c:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:dc616b22757bbd0b1df4e3c1:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.86, + "south": 0.12, + "east": 0.02 + } + }, + { + "id": "snake:dc616b22757bbd0b1df4e3c1:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.71, + "true": 0.29 + } + }, + { + "id": "snake:dc616b22757bbd0b1df4e3c1:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:dc616b22757bbd0b1df4e3c1:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "snake:e81d37b1cc0f0c63ff630476:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.04, + "west": 0.43, + "south": 0.53 + } + }, + { + "id": "snake:e81d37b1cc0f0c63ff630476:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:e81d37b1cc0f0c63ff630476:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "snake:e81d37b1cc0f0c63ff630476:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:2569997e97622b8fb7b847f3:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 186, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.53, + "west": 0.4, + "north": 0.07 + } + }, + { + "id": "snake:2569997e97622b8fb7b847f3:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "snake:2569997e97622b8fb7b847f3:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "snake:2569997e97622b8fb7b847f3:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.34, + "3": 0.43, + "4": 0.06, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 227, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.05, + "south": 0.95 + } + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.49, + "2": 0.32, + "3": 0.04, + "4": 0.01, + "5": 0, + "6": 0.14 + } + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.78, + "true": 0.22 + } + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 227, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.23, + "east": 0.7, + "north": 0.07 + } + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.31, + "3": 0.4, + "4": 0.07, + "5": 0.01, + "6": 0.19 + } + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.69, + "true": 0.31 + } + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.74, + "east": 0.26 + } + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.3, + "3": 0.5700000000000001, + "4": 0.04, + "5": 0, + "6": 0.09 + } + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "snake:7c028b79a76c86db17e4e355:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.01, + "north": 0.02, + "east": 0.97 + } + }, + { + "id": "snake:7c028b79a76c86db17e4e355:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:7c028b79a76c86db17e4e355:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:7c028b79a76c86db17e4e355:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:9a1e13e48c5f3401995d4aea:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.29, + "east": 0, + "south": 0.71 + } + }, + { + "id": "snake:9a1e13e48c5f3401995d4aea:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.84, + "true": 0.16 + } + }, + { + "id": "snake:9a1e13e48c5f3401995d4aea:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:9a1e13e48c5f3401995d4aea:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "snake:f39435f416b3a0bae23647a6:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.87, + "north": 0.12, + "south": 0.01 + } + }, + { + "id": "snake:f39435f416b3a0bae23647a6:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "snake:f39435f416b3a0bae23647a6:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:f39435f416b3a0bae23647a6:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "snake:e85918d93b8f26c1d05dfe14:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.22, + "east": 0.04, + "north": 0.74 + } + }, + { + "id": "snake:e85918d93b8f26c1d05dfe14:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:e85918d93b8f26c1d05dfe14:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "snake:e85918d93b8f26c1d05dfe14:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.15, + "3": 0.55, + "4": 0.11, + "5": 0.01, + "6": 0.18 + } + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.75, + "true": 0.25 + } + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.23, + "north": 0.77 + } + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.17, + "2": 0.48, + "3": 0.13, + "4": 0.01, + "5": 0.01, + "6": 0.2 + } + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8200000000000001, + "true": 0.18 + } + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 224, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.05, + "north": 0.51, + "west": 0.44 + } + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.21, + "3": 0.58, + "4": 0.08, + "5": 0.01, + "6": 0.12 + } + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 221, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.45, + "south": 0.03, + "west": 0.52 + } + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.1, + "2": 0.6, + "3": 0.17, + "4": 0.01, + "5": 0, + "6": 0.12 + } + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "snake:9421d90164ee0257fe3e727d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.65, + "south": 0.01, + "east": 0.34 + } + }, + { + "id": "snake:9421d90164ee0257fe3e727d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:9421d90164ee0257fe3e727d:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:9421d90164ee0257fe3e727d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:d8dda02461c89b190e12b0c9:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.28, + "west": 0.69, + "east": 0.03 + } + }, + { + "id": "snake:d8dda02461c89b190e12b0c9:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "snake:d8dda02461c89b190e12b0c9:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:d8dda02461c89b190e12b0c9:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "snake:64d53b0831fecd4af36767e0:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.06, + "west": 0.33, + "south": 0.61 + } + }, + { + "id": "snake:64d53b0831fecd4af36767e0:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "snake:64d53b0831fecd4af36767e0:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "snake:64d53b0831fecd4af36767e0:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "snake:3664759b6633c11ca5d8e36d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.04, + "west": 0.43, + "north": 0.53 + } + }, + { + "id": "snake:3664759b6633c11ca5d8e36d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:3664759b6633c11ca5d8e36d:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "snake:3664759b6633c11ca5d8e36d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.23, + "north": 0.77 + } + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 189, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 195, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.04, + "3": 0.54, + "4": 0.24, + "5": 0.02, + "6": 0.16 + } + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.9, + "true": 0.1 + } + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.92, + "true": 0.08 + } + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 208, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.94, + "west": 0.01, + "south": 0.05 + } + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 189, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 195, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.08, + "2": 0.45, + "3": 0.23, + "4": 0.06, + "5": 0.01, + "6": 0.17 + } + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.85, + "true": 0.15 + } + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.040000000000000036, + "true": 0.96 + } + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 209, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.09, + "west": 0.16, + "north": 0.75 + } + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 189, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 195, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.08, + "3": 0.61, + "4": 0.17, + "5": 0.02, + "6": 0.12 + } + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.75, + "true": 0.25 + } + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 207, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.08, + "north": 0.75, + "west": 0.17 + } + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 188, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 194, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.07, + "3": 0.66, + "4": 0.15, + "5": 0.01, + "6": 0.11 + } + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8, + "true": 0.2 + } + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:5e7387b379d5ac38a3303d86:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9999999999999999, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.9299999999999999, + "north": 0.04, + "south": 0.03 + } + }, + { + "id": "snake:5e7387b379d5ac38a3303d86:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:5e7387b379d5ac38a3303d86:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:5e7387b379d5ac38a3303d86:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:91c593992ed356c02167928f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.14, + "south": 0.84, + "east": 0.02 + } + }, + { + "id": "snake:91c593992ed356c02167928f:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:91c593992ed356c02167928f:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:91c593992ed356c02167928f:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "snake:b884ef2e8f42ac2d63d357e2:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.03, + "north": 0.32, + "west": 0.65 + } + }, + { + "id": "snake:b884ef2e8f42ac2d63d357e2:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:b884ef2e8f42ac2d63d357e2:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "snake:b884ef2e8f42ac2d63d357e2:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "snake:7deaeb6a65c92f55a9839aee:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.04, + "south": 0.89, + "east": 0.07 + } + }, + { + "id": "snake:7deaeb6a65c92f55a9839aee:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "snake:7deaeb6a65c92f55a9839aee:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "snake:7deaeb6a65c92f55a9839aee:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.03, + "2": 0.28, + "3": 0.37, + "4": 0.09, + "5": 0.01, + "6": 0.22 + } + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.83, + "south": 0.17 + } + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.22, + "3": 0.52, + "4": 0.1, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 221, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0, + "north": 0.05, + "east": 0.95 + } + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.26, + "2": 0.51, + "3": 0.1, + "4": 0.01, + "5": 0, + "6": 0.12 + } + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.82, + "south": 0.18 + } + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.05, + "2": 0.51, + "3": 0.25, + "4": 0.02, + "5": 0, + "6": 0.17 + } + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.72, + "true": 0.28 + } + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:e19f21eb60e12080c813e7d5:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.03, + "east": 0.93, + "north": 0.04 + } + }, + { + "id": "snake:e19f21eb60e12080c813e7d5:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:e19f21eb60e12080c813e7d5:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:e19f21eb60e12080c813e7d5:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:b588a89600945856a65a65d5:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.04, + "south": 0.75, + "west": 0.21 + } + }, + { + "id": "snake:b588a89600945856a65a65d5:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:b588a89600945856a65a65d5:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:b588a89600945856a65a65d5:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "snake:807ed6b55bdbd5cae8404a44:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.36, + "south": 0.02, + "north": 0.62 + } + }, + { + "id": "snake:807ed6b55bdbd5cae8404a44:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "snake:807ed6b55bdbd5cae8404a44:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:807ed6b55bdbd5cae8404a44:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:849d3f6680769a349ea6804e:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.04, + "east": 0.37, + "south": 0.59 + } + }, + { + "id": "snake:849d3f6680769a349ea6804e:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.86, + "true": 0.14 + } + }, + { + "id": "snake:849d3f6680769a349ea6804e:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:849d3f6680769a349ea6804e:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.11, + "3": 0.59, + "4": 0.14, + "5": 0.01, + "6": 0.15 + } + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.63, + "true": 0.37 + } + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6599999999999999, + "true": 0.34 + } + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 228, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.94, + "east": 0.06 + } + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.51, + "2": 0.29, + "3": 0.06, + "4": 0.01, + "5": 0, + "6": 0.13 + } + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.7, + "true": 0.3 + } + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 231, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.48, + "north": 0.49, + "east": 0.03 + } + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.15, + "3": 0.64, + "4": 0.1, + "5": 0.01, + "6": 0.1 + } + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 228, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.12, + "north": 0.88 + } + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.15, + "3": 0.55, + "4": 0.15, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.63, + "true": 0.37 + } + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:96b29694b8f8be77c05a7ad4:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.93, + "north": 0.04, + "south": 0.03 + } + }, + { + "id": "snake:96b29694b8f8be77c05a7ad4:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:96b29694b8f8be77c05a7ad4:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:96b29694b8f8be77c05a7ad4:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:ab770b0e55dc3a2c83839230:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.04, + "north": 0.78, + "east": 0.18 + } + }, + { + "id": "snake:ab770b0e55dc3a2c83839230:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:ab770b0e55dc3a2c83839230:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:ab770b0e55dc3a2c83839230:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:e208c82a1a53aaa466a7d90b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.13, + "west": 0.04, + "north": 0.83 + } + }, + { + "id": "snake:e208c82a1a53aaa466a7d90b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:e208c82a1a53aaa466a7d90b:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "snake:e208c82a1a53aaa466a7d90b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:289128d773eb9f635717f7df:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.01, + "south": 0.56, + "east": 0.43 + } + }, + { + "id": "snake:289128d773eb9f635717f7df:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:289128d773eb9f635717f7df:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:289128d773eb9f635717f7df:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 225, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.03, + "west": 0.97 + } + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.24, + "3": 0.53, + "4": 0.1, + "5": 0.01, + "6": 0.11 + } + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.72, + "true": 0.28 + } + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.64, + "true": 0.36 + } + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.98, + "south": 0.02 + } + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0.01, + "1": 0.8, + "2": 0.02, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.16 + } + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.73, + "true": 0.27 + } + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 226, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.13, + "north": 0.58, + "south": 0.29 + } + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.15, + "3": 0.58, + "4": 0.11, + "5": 0.01, + "6": 0.15 + } + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.47, + "north": 0.53 + } + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.41000000000000003, + "2": 0.32, + "3": 0.12, + "4": 0.02, + "5": 0.01, + "6": 0.12 + } + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.64, + "true": 0.36 + } + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "snake:158c1ec585e67397a4124aef:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.28, + "north": 0.69, + "south": 0.03 + } + }, + { + "id": "snake:158c1ec585e67397a4124aef:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:158c1ec585e67397a4124aef:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:158c1ec585e67397a4124aef:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:3b9af3823f55bc1710f3ffbf:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.02, + "west": 0.35, + "south": 0.63 + } + }, + { + "id": "snake:3b9af3823f55bc1710f3ffbf:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "snake:3b9af3823f55bc1710f3ffbf:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:3b9af3823f55bc1710f3ffbf:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "snake:e7dc5b5e4133d3518dd65a04:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0, + "south": 0, + "west": 1 + } + }, + { + "id": "snake:e7dc5b5e4133d3518dd65a04:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "snake:e7dc5b5e4133d3518dd65a04:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:e7dc5b5e4133d3518dd65a04:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:c2ed8fcd51b7007b0c3d07b5:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "east": 0.05, + "west": 0.56, + "south": 0.38 + } + }, + { + "id": "snake:c2ed8fcd51b7007b0c3d07b5:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "snake:c2ed8fcd51b7007b0c3d07b5:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:c2ed8fcd51b7007b0c3d07b5:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.14, + "north": 0.86 + } + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.43, + "3": 0.29, + "4": 0.11, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.9, + "true": 0.1 + } + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.89, + "true": 0.11 + } + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 218, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0, + "north": 0, + "east": 1 + } + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "0": 0.01, + "1": 0.93, + "2": 0, + "3": 0, + "4": 0, + "5": 0, + "6": 0.05 + } + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 221, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.55, + "west": 0.05, + "east": 0.4 + } + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.11, + "3": 0.58, + "4": 0.17, + "5": 0.01, + "6": 0.13 + } + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.88, + "true": 0.12 + } + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 217, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.01, + "south": 0.05, + "east": 0.78, + "north": 0.16 + } + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.13, + "3": 0.66, + "4": 0.09, + "5": 0.01, + "6": 0.11 + } + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:28659e15e2e3bd051af50b5b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.01, + "east": 0.2, + "north": 0.79 + } + }, + { + "id": "snake:28659e15e2e3bd051af50b5b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:28659e15e2e3bd051af50b5b:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:28659e15e2e3bd051af50b5b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:64cf7df1eefe1400e72f81cb:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.47, + "north": 0.18, + "east": 0.35 + } + }, + { + "id": "snake:64cf7df1eefe1400e72f81cb:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "snake:64cf7df1eefe1400e72f81cb:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "snake:64cf7df1eefe1400e72f81cb:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "snake:83a83973fd86ddcd08278c77:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.47, + "north": 0.11, + "west": 0.42 + } + }, + { + "id": "snake:83a83973fd86ddcd08278c77:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6599999999999999, + "true": 0.34 + } + }, + { + "id": "snake:83a83973fd86ddcd08278c77:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "snake:83a83973fd86ddcd08278c77:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "snake:e8e18f55d8092c0ee71b36a4:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0, + "south": 0.14, + "east": 0.86 + } + }, + { + "id": "snake:e8e18f55d8092c0ee71b36a4:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "snake:e8e18f55d8092c0ee71b36a4:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:e8e18f55d8092c0ee71b36a4:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.03, + "3": 0.25, + "4": 0.43, + "5": 0.13, + "6": 0.16 + } + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 559, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 559, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 559, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.74, + "true": 0.26 + } + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 559, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.21, + "east": 0.79 + } + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.18, + "3": 0.45999999999999996, + "4": 0.19, + "5": 0.02, + "6": 0.14 + } + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 546, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.39, + "west": 0.58, + "east": 0.03 + } + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.02, + "3": 0.32, + "4": 0.4, + "5": 0.11, + "6": 0.15 + } + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.93, + "east": 0.07 + } + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.04, + "3": 0.31, + "4": 0.32, + "5": 0.12, + "6": 0.21 + } + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:587e940982bc37d3a9d3ebbc:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.02, + "east": 0.95, + "south": 0.03 + } + }, + { + "id": "snake:587e940982bc37d3a9d3ebbc:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:587e940982bc37d3a9d3ebbc:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:587e940982bc37d3a9d3ebbc:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:e5222ca460d01b03e75a1acf:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.62, + "east": 0.33, + "north": 0.05 + } + }, + { + "id": "snake:e5222ca460d01b03e75a1acf:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "snake:e5222ca460d01b03e75a1acf:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "snake:e5222ca460d01b03e75a1acf:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "snake:54224971d3a7ebdd73c88f5f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 1, + "south": 0, + "north": 0 + } + }, + { + "id": "snake:54224971d3a7ebdd73c88f5f:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:54224971d3a7ebdd73c88f5f:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:54224971d3a7ebdd73c88f5f:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:548c1d7ff2b216307d82118d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.22, + "north": 0.03, + "south": 0.75 + } + }, + { + "id": "snake:548c1d7ff2b216307d82118d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "snake:548c1d7ff2b216307d82118d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "snake:548c1d7ff2b216307d82118d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 475, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 481, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.29, + "4": 0.43, + "5": 0.13, + "6": 0.14 + } + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 496, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.91, + "east": 0.09 + } + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 476, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 482, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.31, + "3": 0.46, + "4": 0.1, + "5": 0.01, + "6": 0.11 + } + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 498, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.08, + "west": 0.49, + "east": 0.43 + } + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 477, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 483, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.03, + "3": 0.52, + "4": 0.3, + "5": 0.04, + "6": 0.11 + } + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.87, + "west": 0.13 + } + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 474, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 480, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.04, + "3": 0.43, + "4": 0.36, + "5": 0.05, + "6": 0.12 + } + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.71, + "true": 0.29 + } + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "snake:f602fc8e15f8718804a47a1a:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.55, + "north": 0.42, + "south": 0.03 + } + }, + { + "id": "snake:f602fc8e15f8718804a47a1a:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:f602fc8e15f8718804a47a1a:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:f602fc8e15f8718804a47a1a:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:4945a0bf9e0f77491e12fa68:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.27, + "west": 0.7, + "east": 0.03 + } + }, + { + "id": "snake:4945a0bf9e0f77491e12fa68:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:4945a0bf9e0f77491e12fa68:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "snake:4945a0bf9e0f77491e12fa68:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "snake:3c58cce3a6bc409ab2407674:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.97, + "west": 0.01, + "east": 0.02 + } + }, + { + "id": "snake:3c58cce3a6bc409ab2407674:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "snake:3c58cce3a6bc409ab2407674:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:3c58cce3a6bc409ab2407674:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "snake:1fa03fee2824f53457cd8e6d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.07, + "south": 0.03, + "west": 0.9 + } + }, + { + "id": "snake:1fa03fee2824f53457cd8e6d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:1fa03fee2824f53457cd8e6d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "snake:1fa03fee2824f53457cd8e6d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.01, + "west": 0.99 + } + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.03, + "2": 0.63, + "3": 0.15, + "4": 0.04, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 522, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.11, + "east": 0.89 + } + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 502, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.06, + "2": 0.5, + "3": 0.3, + "4": 0.05, + "5": 0.01, + "6": 0.08 + } + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 523, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.62, + "south": 0.02, + "north": 0.36 + } + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 502, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.12, + "3": 0.47, + "4": 0.29, + "5": 0.02, + "6": 0.1 + } + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.74, + "true": 0.26 + } + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 521, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.88, + "north": 0.12 + } + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 501, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.2, + "3": 0.54, + "4": 0.17, + "5": 0.01, + "6": 0.08 + } + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 535, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 535, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 535, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 535, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:21bd326ce12d821777866873:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.8, + "north": 0.19, + "south": 0.01 + } + }, + { + "id": "snake:21bd326ce12d821777866873:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:21bd326ce12d821777866873:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:21bd326ce12d821777866873:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:5ae346163f4ccb67ad5df4cf:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.01, + "east": 0.78, + "north": 0.21 + } + }, + { + "id": "snake:5ae346163f4ccb67ad5df4cf:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:5ae346163f4ccb67ad5df4cf:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "snake:5ae346163f4ccb67ad5df4cf:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "snake:9566442c05492d7ee04c89a0:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 186, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.14, + "north": 0.54, + "east": 0.32 + } + }, + { + "id": "snake:9566442c05492d7ee04c89a0:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:9566442c05492d7ee04c89a0:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "snake:9566442c05492d7ee04c89a0:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "snake:63673902a990ea078c840d71:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 188, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.04, + "north": 0.01, + "west": 0.95 + } + }, + { + "id": "snake:63673902a990ea078c840d71:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:63673902a990ea078c840d71:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "snake:63673902a990ea078c840d71:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.92, + "south": 0.08 + } + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.39, + "4": 0.39, + "5": 0.07, + "6": 0.14 + } + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.88, + "true": 0.12 + } + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.84, + "south": 0.16 + } + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 500, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 506, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.03, + "2": 0.45999999999999996, + "3": 0.37, + "4": 0.05, + "5": 0.01, + "6": 0.08 + } + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.75, + "true": 0.25 + } + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 524, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.41, + "east": 0.09, + "north": 0.5 + } + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.34, + "4": 0.39, + "5": 0.11, + "6": 0.14 + } + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.87, + "true": 0.13 + } + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 525, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.63, + "east": 0.05, + "west": 0.3, + "south": 0.02 + } + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 504, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.02, + "3": 0.47, + "4": 0.37, + "5": 0.04, + "6": 0.1 + } + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 538, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 538, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 538, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 538, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "snake:860d0c0789d6afbac0ce3bba:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.22, + "south": 0.16, + "east": 0.62 + } + }, + { + "id": "snake:860d0c0789d6afbac0ce3bba:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:860d0c0789d6afbac0ce3bba:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:860d0c0789d6afbac0ce3bba:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:775256c2c671ac2e72ebdd24:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.18, + "west": 0.81, + "east": 0.01 + } + }, + { + "id": "snake:775256c2c671ac2e72ebdd24:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "snake:775256c2c671ac2e72ebdd24:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "snake:775256c2c671ac2e72ebdd24:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "snake:38e40a2c70bb7ec9751a2cdf:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.83, + "south": 0.16, + "east": 0.01 + } + }, + { + "id": "snake:38e40a2c70bb7ec9751a2cdf:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "snake:38e40a2c70bb7ec9751a2cdf:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:38e40a2c70bb7ec9751a2cdf:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "snake:4a07bcdf7cf6871e5f690963:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.13, + "north": 0.1, + "west": 0.77 + } + }, + { + "id": "snake:4a07bcdf7cf6871e5f690963:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:4a07bcdf7cf6871e5f690963:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "snake:4a07bcdf7cf6871e5f690963:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 402, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 408, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.06, + "3": 0.46, + "4": 0.24, + "5": 0.04, + "6": 0.2 + } + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 420, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.27, + "north": 0.73 + } + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 400, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 406, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.81, + "2": 0.04, + "3": 0.03, + "4": 0.01, + "5": 0, + "6": 0.11 + } + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 434, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 434, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 434, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 434, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 421, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.14, + "east": 0.84, + "south": 0.02 + } + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 401, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 407, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.06, + "3": 0.56, + "4": 0.2, + "5": 0.02, + "6": 0.15 + } + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 435, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 435, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6599999999999999, + "true": 0.34 + } + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 435, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 435, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 423, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.34, + "north": 0.66 + } + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 402, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 408, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.27, + "3": 0.47, + "4": 0.11, + "5": 0.01, + "6": 0.13 + } + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.64, + "true": 0.36 + } + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "snake:6a67db98c3bc114af9f92771:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "south": 0.55, + "north": 0.05, + "east": 0.39 + } + }, + { + "id": "snake:6a67db98c3bc114af9f92771:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:6a67db98c3bc114af9f92771:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:6a67db98c3bc114af9f92771:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:14fc341d4676da4ba0c2c856:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.65, + "south": 0.07, + "north": 0.28 + } + }, + { + "id": "snake:14fc341d4676da4ba0c2c856:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "snake:14fc341d4676da4ba0c2c856:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:14fc341d4676da4ba0c2c856:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:71ddc8139db3932dc53395ea:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.36, + "west": 0.01, + "east": 0.63 + } + }, + { + "id": "snake:71ddc8139db3932dc53395ea:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "snake:71ddc8139db3932dc53395ea:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:71ddc8139db3932dc53395ea:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "snake:eeeb37680018cbca5c0ce537:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 1, + "north": 0, + "south": 0 + } + }, + { + "id": "snake:eeeb37680018cbca5c0ce537:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:eeeb37680018cbca5c0ce537:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.030000000000000027, + "true": 0.97 + } + }, + { + "id": "snake:eeeb37680018cbca5c0ce537:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 505, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.03, + "3": 0.45, + "4": 0.38, + "5": 0.05, + "6": 0.09 + } + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.21, + "north": 0.79 + } + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 505, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.2, + "3": 0.54, + "4": 0.16, + "5": 0.02, + "6": 0.08 + } + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 517, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.58, + "south": 0.32, + "north": 0.1 + } + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 498, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 504, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.32, + "4": 0.39, + "5": 0.15, + "6": 0.13 + } + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.87, + "east": 0.13 + } + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 500, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 506, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.26, + "4": 0.43, + "5": 0.16, + "6": 0.14 + } + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:16e173e44fd1eb833e530644:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.5700000000000001, + "south": 0.22, + "north": 0.21 + } + }, + { + "id": "snake:16e173e44fd1eb833e530644:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:16e173e44fd1eb833e530644:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:16e173e44fd1eb833e530644:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:ee03d4da232d5808058554c5:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.35, + "west": 0.41, + "north": 0.24 + } + }, + { + "id": "snake:ee03d4da232d5808058554c5:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:ee03d4da232d5808058554c5:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "snake:ee03d4da232d5808058554c5:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:d09b8ff53ace600b84d8a0ad:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.19, + "east": 0.02, + "north": 0.79 + } + }, + { + "id": "snake:d09b8ff53ace600b84d8a0ad:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:d09b8ff53ace600b84d8a0ad:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "snake:d09b8ff53ace600b84d8a0ad:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "snake:6e78fa8aa3855cf9e9a8b2f6:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.02, + "east": 0.44, + "south": 0.54 + } + }, + { + "id": "snake:6e78fa8aa3855cf9e9a8b2f6:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "snake:6e78fa8aa3855cf9e9a8b2f6:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:6e78fa8aa3855cf9e9a8b2f6:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.73, + "true": 0.27 + } + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 517, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.06, + "north": 0.94 + } + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 502, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.38, + "4": 0.41000000000000003, + "5": 0.09, + "6": 0.11 + } + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.81, + "true": 0.19 + } + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 514, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.05, + "north": 0.95 + } + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 495, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 501, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.53, + "2": 0.28, + "3": 0.06, + "4": 0.01, + "5": 0.01, + "6": 0.11 + } + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.75, + "true": 0.25 + } + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 518, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.06, + "south": 0.04, + "north": 0.9 + } + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.02, + "3": 0.31, + "4": 0.43, + "5": 0.12, + "6": 0.12 + } + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 518, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.09, + "north": 0.91 + } + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.3, + "4": 0.45, + "5": 0.11, + "6": 0.13 + } + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "snake:a336313d612f3aa9c3be807b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.01, + "south": 0, + "east": 0.99 + } + }, + { + "id": "snake:a336313d612f3aa9c3be807b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:a336313d612f3aa9c3be807b:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:a336313d612f3aa9c3be807b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:0c81cbbe176f913d0b8e93ea:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.7, + "north": 0.28, + "east": 0.02 + } + }, + { + "id": "snake:0c81cbbe176f913d0b8e93ea:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:0c81cbbe176f913d0b8e93ea:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "snake:0c81cbbe176f913d0b8e93ea:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:6ada38c77274320ad626e980:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.68, + "south": 0.2, + "north": 0.12 + } + }, + { + "id": "snake:6ada38c77274320ad626e980:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "snake:6ada38c77274320ad626e980:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "snake:6ada38c77274320ad626e980:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:403dd328ac9cdb6fc3a2d912:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0, + "west": 0, + "north": 1 + } + }, + { + "id": "snake:403dd328ac9cdb6fc3a2d912:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:403dd328ac9cdb6fc3a2d912:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.85, + "true": 0.15 + } + }, + { + "id": "snake:403dd328ac9cdb6fc3a2d912:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.54, + "north": 0.46 + } + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 491, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.4, + "4": 0.39999999999999997, + "5": 0.09, + "6": 0.1 + } + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.86, + "true": 0.14 + } + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.91, + "true": 0.09 + } + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 512, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.1, + "south": 0.04, + "east": 0.86 + } + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 498, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.82, + "2": 0.07, + "3": 0.03, + "4": 0.01, + "5": 0, + "6": 0.07 + } + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 512, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.05, + "south": 0.02, + "north": 0.52, + "east": 0.41 + } + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 498, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.37, + "4": 0.47, + "5": 0.06, + "6": 0.09 + } + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 510, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.38, + "west": 0.6, + "south": 0.02 + } + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.24, + "3": 0.42, + "4": 0.21, + "5": 0.02, + "6": 0.1 + } + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8200000000000001, + "true": 0.18 + } + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "snake:870e7a8c20e72999a38f1bbb:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.54, + "north": 0.43, + "south": 0.03 + } + }, + { + "id": "snake:870e7a8c20e72999a38f1bbb:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:870e7a8c20e72999a38f1bbb:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:870e7a8c20e72999a38f1bbb:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:761db75641266f665d423dd6:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0, + "west": 0.98, + "north": 0.02 + } + }, + { + "id": "snake:761db75641266f665d423dd6:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:761db75641266f665d423dd6:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:761db75641266f665d423dd6:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:f3d23931d79320f7dcb6b809:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.34, + "south": 0.58, + "north": 0.08 + } + }, + { + "id": "snake:f3d23931d79320f7dcb6b809:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:f3d23931d79320f7dcb6b809:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "snake:f3d23931d79320f7dcb6b809:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:733f1c4c8d36eef7b6a3a628:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.29, + "east": 0.66, + "south": 0.05 + } + }, + { + "id": "snake:733f1c4c8d36eef7b6a3a628:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "snake:733f1c4c8d36eef7b6a3a628:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:733f1c4c8d36eef7b6a3a628:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 399, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 405, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.03, + "3": 0.33, + "4": 0.38, + "5": 0.08, + "6": 0.18 + } + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 419, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.11, + "east": 0.89 + } + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 399, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 405, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.36, + "2": 0.39, + "3": 0.08, + "4": 0.02, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 420, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.53, + "south": 0.04, + "east": 0.43 + } + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 399, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 405, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.04, + "3": 0.4, + "4": 0.33, + "5": 0.06, + "6": 0.17 + } + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 418, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.19, + "east": 0.81 + } + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 398, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 404, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.15, + "3": 0.51, + "4": 0.16, + "5": 0.02, + "6": 0.16 + } + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 432, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 432, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 432, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 432, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:58998e2b30200567009458f2:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.02, + "east": 0.61, + "north": 0.37 + } + }, + { + "id": "snake:58998e2b30200567009458f2:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:58998e2b30200567009458f2:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:58998e2b30200567009458f2:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:d20444edd7bce4c96411c22d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 1, + "south": 0, + "north": 0 + } + }, + { + "id": "snake:d20444edd7bce4c96411c22d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:d20444edd7bce4c96411c22d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:d20444edd7bce4c96411c22d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:e9f2b3223460fa3d568ffc9f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.44, + "north": 0.55, + "east": 0.01 + } + }, + { + "id": "snake:e9f2b3223460fa3d568ffc9f:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:e9f2b3223460fa3d568ffc9f:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "snake:e9f2b3223460fa3d568ffc9f:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:621b671d54f22d05c2855b2b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.69, + "west": 0.28, + "north": 0.03 + } + }, + { + "id": "snake:621b671d54f22d05c2855b2b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:621b671d54f22d05c2855b2b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "snake:621b671d54f22d05c2855b2b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.36, + "3": 0.45, + "4": 0.09, + "5": 0.01, + "6": 0.08 + } + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.9, + "west": 0.1 + } + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.51, + "3": 0.34, + "4": 0.05, + "5": 0.01, + "6": 0.07 + } + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 521, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 521, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 521, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 521, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 507, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.61, + "east": 0.11, + "north": 0.28 + } + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.05, + "2": 0.57, + "3": 0.22, + "4": 0.04, + "5": 0.01, + "6": 0.11 + } + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.7, + "true": 0.3 + } + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 506, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.04, + "east": 0.96 + } + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.77, + "2": 0.09, + "3": 0.03, + "4": 0.01, + "5": 0, + "6": 0.1 + } + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.79, + "true": 0.21 + } + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "snake:a90f25174a8a6744aac7d061:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.73, + "south": 0.01, + "north": 0.26 + } + }, + { + "id": "snake:a90f25174a8a6744aac7d061:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:a90f25174a8a6744aac7d061:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:a90f25174a8a6744aac7d061:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:3a60ced263839ebe4ffa236f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.01, + "north": 0.02, + "west": 0.97 + } + }, + { + "id": "snake:3a60ced263839ebe4ffa236f:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:3a60ced263839ebe4ffa236f:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:3a60ced263839ebe4ffa236f:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:5566ec4180230f4b928c1c51:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0, + "south": 1, + "east": 0 + } + }, + { + "id": "snake:5566ec4180230f4b928c1c51:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "snake:5566ec4180230f4b928c1c51:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:5566ec4180230f4b928c1c51:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "snake:17d6eed389e537292d04e135:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 187, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.19, + "north": 0.59, + "east": 0.22 + } + }, + { + "id": "snake:17d6eed389e537292d04e135:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:17d6eed389e537292d04e135:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "snake:17d6eed389e537292d04e135:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 516, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.87, + "north": 0.13 + } + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.3, + "4": 0.42, + "5": 0.14, + "6": 0.13 + } + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.86, + "true": 0.14 + } + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.95, + "north": 0.05 + } + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 498, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 504, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.12, + "2": 0.56, + "3": 0.13, + "4": 0.04, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.85, + "true": 0.15 + } + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 518, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.01, + "north": 0.05, + "east": 0.94 + } + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 498, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 504, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.05, + "3": 0.59, + "4": 0.19, + "5": 0.05, + "6": 0.12 + } + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 517, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.13, + "south": 0.87 + } + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.34, + "4": 0.42, + "5": 0.12, + "6": 0.11 + } + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "snake:e6fc77aaa094d0d78957065e:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.49, + "south": 0.23, + "north": 0.28 + } + }, + { + "id": "snake:e6fc77aaa094d0d78957065e:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:e6fc77aaa094d0d78957065e:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:e6fc77aaa094d0d78957065e:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:7b1c6353934e5e0b84835f97:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.02, + "north": 0.79, + "east": 0.19 + } + }, + { + "id": "snake:7b1c6353934e5e0b84835f97:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:7b1c6353934e5e0b84835f97:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:7b1c6353934e5e0b84835f97:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:69fefba3afa3e52436b6a3d9:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.66, + "west": 0.23, + "east": 0.11 + } + }, + { + "id": "snake:69fefba3afa3e52436b6a3d9:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:69fefba3afa3e52436b6a3d9:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:69fefba3afa3e52436b6a3d9:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 476, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 482, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.36, + "4": 0.37, + "5": 0.09, + "6": 0.17 + } + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.89, + "true": 0.11 + } + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.92, + "true": 0.08 + } + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 497, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.41, + "west": 0.1, + "north": 0.49 + } + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 477, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 483, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.19, + "3": 0.53, + "4": 0.17, + "5": 0.01, + "6": 0.1 + } + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 496, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.07, + "east": 0.45, + "north": 0.45, + "south": 0.03 + } + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 477, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 483, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.02, + "3": 0.46, + "4": 0.3, + "5": 0.07, + "6": 0.15 + } + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 495, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.91, + "south": 0.07, + "west": 0.02 + } + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 476, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 482, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.05, + "3": 0.47, + "4": 0.23, + "5": 0.08, + "6": 0.16 + } + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.83, + "true": 0.17 + } + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:11f37c8a34e52be1e5290999:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.77, + "north": 0.22, + "south": 0.01 + } + }, + { + "id": "snake:11f37c8a34e52be1e5290999:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:11f37c8a34e52be1e5290999:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:11f37c8a34e52be1e5290999:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:fda49c37ac5bc4d0ace320fa:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.01, + "east": 0.35, + "south": 0.64 + } + }, + { + "id": "snake:fda49c37ac5bc4d0ace320fa:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "snake:fda49c37ac5bc4d0ace320fa:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:fda49c37ac5bc4d0ace320fa:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "snake:389deeada306290bf795e0c1:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.01, + "west": 0.18, + "north": 0.81 + } + }, + { + "id": "snake:389deeada306290bf795e0c1:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "snake:389deeada306290bf795e0c1:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "snake:389deeada306290bf795e0c1:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:01ae3faf9d37bf799bde6b7b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.07, + "west": 0.92, + "east": 0.01 + } + }, + { + "id": "snake:01ae3faf9d37bf799bde6b7b:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:01ae3faf9d37bf799bde6b7b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:01ae3faf9d37bf799bde6b7b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.41, + "2": 0.27, + "3": 0.03, + "4": 0.01, + "5": 0.01, + "6": 0.27 + } + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.72, + "true": 0.28 + } + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.83, + "east": 0.17 + } + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.58, + "2": 0.22, + "3": 0.03, + "4": 0.01, + "5": 0.01, + "6": 0.15 + } + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.64, + "true": 0.36 + } + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.96, + "south": 0.04 + } + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.48, + "2": 0.23, + "3": 0.04, + "4": 0.01, + "5": 0.01, + "6": 0.23 + } + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.73, + "true": 0.27 + } + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.78, + "true": 0.22 + } + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.7, + "north": 0.3 + } + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.13, + "2": 0.53, + "3": 0.09, + "4": 0.01, + "5": 0.01, + "6": 0.23 + } + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6699999999999999, + "true": 0.33 + } + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:a8b7df91259c3f05aeb382b4:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.23, + "north": 0.02, + "south": 0.75 + } + }, + { + "id": "snake:a8b7df91259c3f05aeb382b4:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "snake:a8b7df91259c3f05aeb382b4:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "snake:a8b7df91259c3f05aeb382b4:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:df90ae5ad4e7c29cddbeece3:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.8099999999999999, + "east": 0.12, + "west": 0.07 + } + }, + { + "id": "snake:df90ae5ad4e7c29cddbeece3:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "snake:df90ae5ad4e7c29cddbeece3:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:df90ae5ad4e7c29cddbeece3:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "snake:cd94f508a6fac43a109bbad1:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.89, + "west": 0.02, + "east": 0.09 + } + }, + { + "id": "snake:cd94f508a6fac43a109bbad1:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:cd94f508a6fac43a109bbad1:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "snake:cd94f508a6fac43a109bbad1:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.36, + "2": 0.33, + "3": 0.04, + "4": 0.01, + "5": 0.01, + "6": 0.25 + } + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0, + "west": 1 + } + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "0": 0, + "1": 0.8, + "2": 0.05, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.13 + } + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.73, + "true": 0.27 + } + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.93, + "east": 0.07 + } + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.49, + "2": 0.27, + "3": 0.02, + "4": 0.01, + "5": 0, + "6": 0.21 + } + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.7, + "true": 0.3 + } + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6599999999999999, + "true": 0.34 + } + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.24, + "east": 0.76 + } + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.74, + "2": 0.09, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.16 + } + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.64, + "true": 0.36 + } + }, + { + "id": "snake:c388fa806f00892833d237ff:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.67, + "east": 0.31, + "south": 0.02 + } + }, + { + "id": "snake:c388fa806f00892833d237ff:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:c388fa806f00892833d237ff:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:c388fa806f00892833d237ff:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:4c5c5ed5f1c363a3abb500d2:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.84, + "south": 0.14, + "north": 0.02 + } + }, + { + "id": "snake:4c5c5ed5f1c363a3abb500d2:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:4c5c5ed5f1c363a3abb500d2:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:4c5c5ed5f1c363a3abb500d2:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:576e0861df9167b47f30ee2e:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.52, + "east": 0.06, + "west": 0.42 + } + }, + { + "id": "snake:576e0861df9167b47f30ee2e:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "snake:576e0861df9167b47f30ee2e:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "snake:576e0861df9167b47f30ee2e:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "snake:52fbbba9d2809da41935b256:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.01, + "south": 0.01, + "west": 0.98 + } + }, + { + "id": "snake:52fbbba9d2809da41935b256:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "snake:52fbbba9d2809da41935b256:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "snake:52fbbba9d2809da41935b256:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.15, + "2": 0.6, + "3": 0.06, + "4": 0.01, + "5": 0.01, + "6": 0.17 + } + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.69, + "true": 0.31 + } + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.88, + "east": 0.12 + } + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.85, + "2": 0.05, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.09 + } + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6599999999999999, + "true": 0.34 + } + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.7, + "true": 0.3 + } + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.95, + "south": 0.01, + "north": 0.04 + } + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "0": 0, + "1": 0.56, + "2": 0.26, + "3": 0.03, + "4": 0, + "5": 0, + "6": 0.14 + } + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.75, + "true": 0.25 + } + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.78, + "east": 0.22 + } + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.77, + "2": 0.06, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.16 + } + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "snake:67cc2d1e520f345a77c25258:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.17, + "south": 0.01, + "east": 0.82 + } + }, + { + "id": "snake:67cc2d1e520f345a77c25258:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:67cc2d1e520f345a77c25258:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:67cc2d1e520f345a77c25258:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:4e0dd4fb1004b1b5f10505fb:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.1, + "south": 0.63, + "east": 0.27 + } + }, + { + "id": "snake:4e0dd4fb1004b1b5f10505fb:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:4e0dd4fb1004b1b5f10505fb:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:4e0dd4fb1004b1b5f10505fb:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "snake:a156b999ce6a51e078e7d63b:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.52, + "east": 0.46, + "south": 0.02 + } + }, + { + "id": "snake:a156b999ce6a51e078e7d63b:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "snake:a156b999ce6a51e078e7d63b:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:a156b999ce6a51e078e7d63b:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "snake:9471f04d7505354571f2afae:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.25, + "south": 0.73, + "north": 0.02 + } + }, + { + "id": "snake:9471f04d7505354571f2afae:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:9471f04d7505354571f2afae:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "snake:9471f04d7505354571f2afae:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.54, + "3": 0.38, + "4": 0.01, + "5": 0, + "6": 0.07 + } + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.89, + "true": 0.11 + } + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8200000000000001, + "true": 0.18 + } + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.38, + "east": 0.02, + "west": 0.59, + "south": 0.01 + } + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.47000000000000003, + "2": 0.44, + "3": 0.03, + "4": 0, + "5": 0, + "6": 0.06 + } + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.04, + "west": 0.85, + "south": 0.11 + } + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.6, + "3": 0.3, + "4": 0.01, + "5": 0, + "6": 0.08 + } + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.01, + "east": 0.04, + "north": 0.58, + "west": 0.37 + } + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.29, + "2": 0.59, + "3": 0.05, + "4": 0, + "5": 0, + "6": 0.07 + } + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "snake:78b599638feb2b5738052f42:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.69, + "north": 0.07, + "south": 0.24 + } + }, + { + "id": "snake:78b599638feb2b5738052f42:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:78b599638feb2b5738052f42:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:78b599638feb2b5738052f42:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:d2e3d53861d338d18932e5c3:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.39, + "north": 0.19, + "west": 0.42 + } + }, + { + "id": "snake:d2e3d53861d338d18932e5c3:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:d2e3d53861d338d18932e5c3:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "snake:d2e3d53861d338d18932e5c3:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:655d2c7cb308d413d245f479:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.47, + "west": 0.51, + "east": 0.02 + } + }, + { + "id": "snake:655d2c7cb308d413d245f479:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "snake:655d2c7cb308d413d245f479:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:655d2c7cb308d413d245f479:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "snake:e1714aca03074946552e86d7:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.38, + "east": 0.01, + "south": 0.61 + } + }, + { + "id": "snake:e1714aca03074946552e86d7:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.78, + "true": 0.22 + } + }, + { + "id": "snake:e1714aca03074946552e86d7:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:e1714aca03074946552e86d7:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 188, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 194, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.55, + "2": 0.12, + "3": 0.03, + "4": 0.01, + "5": 0, + "6": 0.29 + } + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.78, + "true": 0.22 + } + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.17, + "east": 0.83 + } + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 190, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 196, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.07, + "2": 0.63, + "3": 0.19, + "4": 0.02, + "5": 0, + "6": 0.09 + } + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 224, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 224, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6599999999999999, + "true": 0.34 + } + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 224, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 224, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 208, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.23, + "west": 0.72, + "north": 0.05 + } + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 188, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 194, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.1, + "2": 0.59, + "3": 0.19, + "4": 0.02, + "5": 0, + "6": 0.1 + } + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.84, + "true": 0.16 + } + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.81, + "east": 0.19 + } + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 189, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 195, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.07, + "2": 0.59, + "3": 0.19, + "4": 0.02, + "5": 0.01, + "6": 0.12 + } + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "snake:7f5e2225e29d3de69626ddb4:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.33, + "north": 0.64, + "south": 0.03 + } + }, + { + "id": "snake:7f5e2225e29d3de69626ddb4:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:7f5e2225e29d3de69626ddb4:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:7f5e2225e29d3de69626ddb4:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:ca33cb36c8613e8697afb773:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.21, + "north": 0.02, + "east": 0.77 + } + }, + { + "id": "snake:ca33cb36c8613e8697afb773:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8200000000000001, + "true": 0.18 + } + }, + { + "id": "snake:ca33cb36c8613e8697afb773:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:ca33cb36c8613e8697afb773:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "snake:b7664a71002acbd88104c520:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.11, + "south": 0.34, + "east": 0.55 + } + }, + { + "id": "snake:b7664a71002acbd88104c520:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.83, + "true": 0.17 + } + }, + { + "id": "snake:b7664a71002acbd88104c520:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "snake:b7664a71002acbd88104c520:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "snake:57d32914072132dfcd28dde9:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.01, + "west": 0.07, + "north": 0.92 + } + }, + { + "id": "snake:57d32914072132dfcd28dde9:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:57d32914072132dfcd28dde9:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:57d32914072132dfcd28dde9:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.45, + "3": 0.35, + "4": 0.03, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 228, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.03, + "south": 0.23, + "west": 0.74 + } + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.14, + "2": 0.58, + "3": 0.14, + "4": 0.02, + "5": 0.01, + "6": 0.11 + } + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 229, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.08, + "west": 0.79, + "south": 0.13 + } + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.18, + "3": 0.58, + "4": 0.08, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.81, + "true": 0.19 + } + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.7, + "south": 0.3 + } + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.16, + "3": 0.55, + "4": 0.1, + "5": 0.01, + "6": 0.17 + } + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "snake:c8567147711d9e316a6aa4cf:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.25, + "south": 0.09, + "east": 0.66 + } + }, + { + "id": "snake:c8567147711d9e316a6aa4cf:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:c8567147711d9e316a6aa4cf:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:c8567147711d9e316a6aa4cf:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:9e09d6920ac6a46a73bb0a67:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 177, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.28, + "east": 0.51, + "north": 0.21 + } + }, + { + "id": "snake:9e09d6920ac6a46a73bb0a67:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "snake:9e09d6920ac6a46a73bb0a67:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:9e09d6920ac6a46a73bb0a67:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "snake:432bf9e1b22aee6ca9a73089:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.78, + "west": 0.11, + "south": 0.11 + } + }, + { + "id": "snake:432bf9e1b22aee6ca9a73089:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "snake:432bf9e1b22aee6ca9a73089:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "snake:432bf9e1b22aee6ca9a73089:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "snake:6d101d00264544fd0e797201:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.15, + "south": 0.69, + "west": 0.16 + } + }, + { + "id": "snake:6d101d00264544fd0e797201:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:6d101d00264544fd0e797201:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "snake:6d101d00264544fd0e797201:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 196, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.73, + "2": 0.07, + "3": 0.03, + "4": 0.01, + "5": 0, + "6": 0.16 + } + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.64, + "true": 0.36 + } + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6699999999999999, + "true": 0.33 + } + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.82, + "east": 0.18 + } + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 196, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.47, + "2": 0.29, + "3": 0.06, + "4": 0.01, + "5": 0, + "6": 0.17 + } + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.64, + "true": 0.36 + } + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 218, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.54, + "south": 0.05, + "east": 0.41 + } + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.03, + "2": 0.35, + "3": 0.34, + "4": 0.08, + "5": 0.01, + "6": 0.19 + } + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 217, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.72, + "north": 0.23, + "south": 0.05 + } + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 197, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.26, + "3": 0.52, + "4": 0.06, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "snake:0832e3f7e20036db9aa11da9:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.72, + "south": 0.12, + "north": 0.16 + } + }, + { + "id": "snake:0832e3f7e20036db9aa11da9:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:0832e3f7e20036db9aa11da9:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:0832e3f7e20036db9aa11da9:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:994f831336c3c300105f9539:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.2, + "north": 0.8, + "west": 0 + } + }, + { + "id": "snake:994f831336c3c300105f9539:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "snake:994f831336c3c300105f9539:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:994f831336c3c300105f9539:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:77649ea6ad51f9563a36480c:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.6, + "south": 0.21, + "north": 0.19 + } + }, + { + "id": "snake:77649ea6ad51f9563a36480c:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:77649ea6ad51f9563a36480c:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "snake:77649ea6ad51f9563a36480c:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "snake:148cdef4390f40347ee09888:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.45, + "north": 0.02, + "south": 0.53 + } + }, + { + "id": "snake:148cdef4390f40347ee09888:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.86, + "true": 0.14 + } + }, + { + "id": "snake:148cdef4390f40347ee09888:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:148cdef4390f40347ee09888:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 227, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.4, + "3": 0.45, + "4": 0.03, + "5": 0, + "6": 0.11 + } + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8200000000000001, + "true": 0.18 + } + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0, + "east": 1 + } + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 229, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.97, + "2": 0.01, + "3": 0, + "4": 0, + "5": 0, + "6": 0.02 + } + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 257, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 257, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.9, + "true": 0.1 + } + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 257, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.040000000000000036, + "true": 0.96 + } + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 257, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 242, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.05, + "east": 0.84, + "north": 0.11 + } + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 228, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.41, + "3": 0.36, + "4": 0.07, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 256, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 256, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.91, + "true": 0.09 + } + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 256, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 256, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 240, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.04, + "east": 0.63, + "north": 0.04, + "south": 0.29 + } + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 227, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.66, + "4": 0.24, + "5": 0.01, + "6": 0.08 + } + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:8f71ea5f3a25834336bcc05d:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.78, + "north": 0.13, + "south": 0.09 + } + }, + { + "id": "snake:8f71ea5f3a25834336bcc05d:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:8f71ea5f3a25834336bcc05d:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:8f71ea5f3a25834336bcc05d:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:6ac6bd536b0757091c5455f0:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.13, + "south": 0.85, + "west": 0.02 + } + }, + { + "id": "snake:6ac6bd536b0757091c5455f0:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "snake:6ac6bd536b0757091c5455f0:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:6ac6bd536b0757091c5455f0:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "snake:964476c58e1e5730badc9744:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.87, + "west": 0.07, + "east": 0.06 + } + }, + { + "id": "snake:964476c58e1e5730badc9744:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:964476c58e1e5730badc9744:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "snake:964476c58e1e5730badc9744:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "snake:7a57be18ed7438683c73015f:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.12, + "west": 0.86, + "north": 0.02 + } + }, + { + "id": "snake:7a57be18ed7438683c73015f:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.81, + "true": 0.19 + } + }, + { + "id": "snake:7a57be18ed7438683c73015f:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "snake:7a57be18ed7438683c73015f:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 423, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.04, + "3": 0.38, + "4": 0.34, + "5": 0.07, + "6": 0.17 + } + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 434, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.71, + "north": 0.29 + } + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 415, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 421, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.32, + "4": 0.4, + "5": 0.11, + "6": 0.16 + } + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 449, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 449, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 449, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 449, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 438, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.13, + "south": 0.46, + "east": 0.41 + } + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 423, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.35000000000000003, + "4": 0.32, + "5": 0.14, + "6": 0.18 + } + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.11, + "south": 0.89 + } + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 422, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.02, + "3": 0.34, + "4": 0.37, + "5": 0.1, + "6": 0.17 + } + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 450, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 450, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 450, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 450, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "snake:051e6e8a2266fa53121d8477:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.2, + "south": 0.01, + "east": 0.79 + } + }, + { + "id": "snake:051e6e8a2266fa53121d8477:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:051e6e8a2266fa53121d8477:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:051e6e8a2266fa53121d8477:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:0f28d07901bcbe13b678ee02:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.08, + "east": 0.91, + "north": 0.01 + } + }, + { + "id": "snake:0f28d07901bcbe13b678ee02:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:0f28d07901bcbe13b678ee02:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:0f28d07901bcbe13b678ee02:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:e7ba57b10725beea1c1e6143:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.07, + "west": 0.32, + "south": 0.61 + } + }, + { + "id": "snake:e7ba57b10725beea1c1e6143:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:e7ba57b10725beea1c1e6143:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:e7ba57b10725beea1c1e6143:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "snake:3d0675a33c07627aadd5584d:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.01, + "south": 0, + "west": 0.99 + } + }, + { + "id": "snake:3d0675a33c07627aadd5584d:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:3d0675a33c07627aadd5584d:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.9, + "true": 0.1 + } + }, + { + "id": "snake:3d0675a33c07627aadd5584d:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 513, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.14, + "3": 0.48, + "4": 0.25, + "5": 0.02, + "6": 0.11 + } + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.44, + "east": 0.56 + } + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 513, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.23, + "3": 0.53, + "4": 0.13, + "5": 0.01, + "6": 0.09 + } + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 529, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.13, + "north": 0.12, + "west": 0.75 + } + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 514, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.05, + "3": 0.38, + "4": 0.39, + "5": 0.06, + "6": 0.12 + } + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.94, + "east": 0.06 + } + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 514, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.04, + "3": 0.33, + "4": 0.41, + "5": 0.11, + "6": 0.11 + } + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "snake:883677764232a4c9ad39f564:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.92, + "south": 0.05, + "north": 0.03 + } + }, + { + "id": "snake:883677764232a4c9ad39f564:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:883677764232a4c9ad39f564:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:883677764232a4c9ad39f564:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:4ce2ba2de3d075d05fbf6931:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.64, + "east": 0.08, + "west": 0.28 + } + }, + { + "id": "snake:4ce2ba2de3d075d05fbf6931:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:4ce2ba2de3d075d05fbf6931:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:4ce2ba2de3d075d05fbf6931:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "snake:49bd5f26505a7ebdca09e2e6:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.31, + "south": 0.62, + "east": 0.07 + } + }, + { + "id": "snake:49bd5f26505a7ebdca09e2e6:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "snake:49bd5f26505a7ebdca09e2e6:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:49bd5f26505a7ebdca09e2e6:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "snake:8bb9604a0be751b4ad25ceae:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.1, + "west": 0.79, + "north": 0.11 + } + }, + { + "id": "snake:8bb9604a0be751b4ad25ceae:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "snake:8bb9604a0be751b4ad25ceae:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "snake:8bb9604a0be751b4ad25ceae:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 500, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.02, + "3": 0.3, + "4": 0.4, + "5": 0.15, + "6": 0.13 + } + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.63, + "true": 0.37 + } + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.73, + "true": 0.27 + } + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 515, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.92, + "west": 0.08 + } + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 495, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 501, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.54, + "3": 0.27, + "4": 0.04, + "5": 0.01, + "6": 0.12 + } + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 517, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.41, + "west": 0.5, + "south": 0.09 + } + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 502, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.02, + "3": 0.42, + "4": 0.33, + "5": 0.08, + "6": 0.15 + } + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.7, + "true": 0.3 + } + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 514, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.03, + "east": 0.97 + } + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 500, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.03, + "2": 0.63, + "3": 0.15, + "4": 0.05, + "5": 0.01, + "6": 0.13 + } + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.72, + "true": 0.28 + } + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "snake:c0df79c78d46b682825412f6:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.76, + "north": 0.23, + "south": 0.01 + } + }, + { + "id": "snake:c0df79c78d46b682825412f6:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:c0df79c78d46b682825412f6:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:c0df79c78d46b682825412f6:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:11d45c3c68b9900a8ac18e46:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.18, + "east": 0.03, + "west": 0.79 + } + }, + { + "id": "snake:11d45c3c68b9900a8ac18e46:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:11d45c3c68b9900a8ac18e46:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:11d45c3c68b9900a8ac18e46:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:085a2608e9718db6826e4fcd:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.64, + "north": 0.32, + "east": 0.04 + } + }, + { + "id": "snake:085a2608e9718db6826e4fcd:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:085a2608e9718db6826e4fcd:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "snake:085a2608e9718db6826e4fcd:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 344, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 350, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.35, + "4": 0.42, + "5": 0.07, + "6": 0.15 + } + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.81, + "true": 0.19 + } + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.89, + "true": 0.11 + } + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 367, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.09, + "east": 0.91 + } + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 346, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 352, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.43, + "3": 0.4, + "4": 0.05, + "5": 0.01, + "6": 0.09 + } + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 380, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 380, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 380, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 380, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 364, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.68, + "north": 0.08, + "east": 0.24 + } + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 344, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 350, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.25, + "4": 0.41, + "5": 0.16, + "6": 0.17 + } + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 368, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.42, + "west": 0.05, + "east": 0.53 + } + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 347, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 353, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.25, + "3": 0.49, + "4": 0.13, + "5": 0.01, + "6": 0.11 + } + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 381, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 381, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 381, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 381, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "snake:092d50b97340fa17002cb4ab:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.63, + "south": 0.34, + "east": 0.03 + } + }, + { + "id": "snake:092d50b97340fa17002cb4ab:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "snake:092d50b97340fa17002cb4ab:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:092d50b97340fa17002cb4ab:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:417fceca334f03fa9927ab18:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0, + "south": 0.93, + "east": 0.07 + } + }, + { + "id": "snake:417fceca334f03fa9927ab18:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "snake:417fceca334f03fa9927ab18:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:417fceca334f03fa9927ab18:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.92, + "true": 0.08 + } + }, + { + "id": "snake:55c505682323e7ba8a0a15c0:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.03, + "east": 0.47, + "south": 0.5 + } + }, + { + "id": "snake:55c505682323e7ba8a0a15c0:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "snake:55c505682323e7ba8a0a15c0:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:55c505682323e7ba8a0a15c0:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.35, + "2": 0.42, + "3": 0.02, + "4": 0.01, + "5": 0.01, + "6": 0.19 + } + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8, + "true": 0.2 + } + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.97, + "east": 0.03 + } + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.58, + "2": 0.25, + "3": 0.02, + "4": 0, + "5": 0, + "6": 0.15 + } + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.39, + "north": 0.61 + } + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.54, + "2": 0.07, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.38 + } + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.79, + "true": 0.21 + } + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.62, + "south": 0.38 + } + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.68, + "2": 0.1, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.21 + } + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8, + "true": 0.2 + } + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "snake:75a4f70d771aa4b61f9c0d3b:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.07, + "east": 0.9, + "north": 0.03 + } + }, + { + "id": "snake:75a4f70d771aa4b61f9c0d3b:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:75a4f70d771aa4b61f9c0d3b:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:75a4f70d771aa4b61f9c0d3b:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:01ff75816b735bcdd1aeda73:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.06, + "north": 0.21, + "west": 0.73 + } + }, + { + "id": "snake:01ff75816b735bcdd1aeda73:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "snake:01ff75816b735bcdd1aeda73:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.73, + "true": 0.27 + } + }, + { + "id": "snake:01ff75816b735bcdd1aeda73:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "snake:9261d348d2b3d6d8b0a34f81:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.74, + "north": 0.02, + "south": 0.24 + } + }, + { + "id": "snake:9261d348d2b3d6d8b0a34f81:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.7, + "true": 0.3 + } + }, + { + "id": "snake:9261d348d2b3d6d8b0a34f81:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "snake:9261d348d2b3d6d8b0a34f81:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:d8903c26d0c9b4585993dd4f:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 188, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.72, + "north": 0.23, + "south": 0.05 + } + }, + { + "id": "snake:d8903c26d0c9b4585993dd4f:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "snake:d8903c26d0c9b4585993dd4f:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "snake:d8903c26d0c9b4585993dd4f:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.35, + "2": 0.17, + "3": 0.02, + "4": 0, + "5": 0, + "6": 0.46 + } + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6599999999999999, + "true": 0.34 + } + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.81, + "true": 0.19 + } + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.84, + "north": 0.02, + "west": 0.14 + } + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.59, + "2": 0.16, + "3": 0.02, + "4": 0, + "5": 0, + "6": 0.23 + } + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.85, + "true": 0.15 + } + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 151, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.12, + "east": 0.88 + } + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 132, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.42, + "2": 0.35, + "3": 0.05, + "4": 0.01, + "5": 0, + "6": 0.17 + } + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.84, + "west": 0.16 + } + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.55, + "2": 0.25, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.19 + } + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.7, + "true": 0.3 + } + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "snake:0cde1d49787ade977529f5d1:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.83, + "east": 0.01, + "west": 0.16 + } + }, + { + "id": "snake:0cde1d49787ade977529f5d1:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:0cde1d49787ade977529f5d1:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.81, + "true": 0.19 + } + }, + { + "id": "snake:0cde1d49787ade977529f5d1:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "snake:58a227e097f6ef3b33076919:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.01, + "east": 0.74, + "north": 0.25 + } + }, + { + "id": "snake:58a227e097f6ef3b33076919:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:58a227e097f6ef3b33076919:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "snake:58a227e097f6ef3b33076919:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "snake:08b1f28ba1a403013512cf46:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.24, + "north": 0.02, + "south": 0.74 + } + }, + { + "id": "snake:08b1f28ba1a403013512cf46:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.69, + "true": 0.31 + } + }, + { + "id": "snake:08b1f28ba1a403013512cf46:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:08b1f28ba1a403013512cf46:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.59, + "2": 0.05, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.35 + } + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.75, + "true": 0.25 + } + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.74, + "true": 0.26 + } + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.91, + "south": 0.09 + } + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.67, + "2": 0.16, + "3": 0.01, + "4": 0, + "5": 0, + "6": 0.16 + } + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.79, + "true": 0.21 + } + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.02, + "north": 0.51, + "west": 0.47 + } + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.47, + "2": 0.33, + "3": 0.03, + "4": 0.01, + "5": 0, + "6": 0.16 + } + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 1, + "south": 0 + } + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.92, + "2": 0.01, + "3": 0, + "4": 0, + "5": 0, + "6": 0.07 + } + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "snake:abb1527dca73ee5804fa14e9:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.78, + "south": 0.2, + "east": 0.02 + } + }, + { + "id": "snake:abb1527dca73ee5804fa14e9:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:abb1527dca73ee5804fa14e9:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:abb1527dca73ee5804fa14e9:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "snake:070695834c3b294ad333f6ec:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.21, + "south": 0.02, + "north": 0.77 + } + }, + { + "id": "snake:070695834c3b294ad333f6ec:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:070695834c3b294ad333f6ec:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:070695834c3b294ad333f6ec:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "snake:f8db42503df5695a2a3ed03d:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.53, + "south": 0.15, + "north": 0.32 + } + }, + { + "id": "snake:f8db42503df5695a2a3ed03d:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "snake:f8db42503df5695a2a3ed03d:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:f8db42503df5695a2a3ed03d:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.08, + "2": 0.62, + "3": 0.12, + "4": 0.01, + "5": 0.01, + "6": 0.16 + } + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.83, + "true": 0.17 + } + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.86, + "north": 0.13, + "east": 0.01 + } + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.71, + "2": 0.13, + "3": 0.04, + "4": 0.01, + "5": 0, + "6": 0.11 + } + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.81, + "true": 0.19 + } + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.56, + "east": 0.41, + "south": 0.03 + } + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.6, + "3": 0.18, + "4": 0.01, + "5": 0.01, + "6": 0.18 + } + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.7, + "true": 0.3 + } + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.74, + "south": 0.26 + } + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.69, + "2": 0.23, + "3": 0.03, + "4": 0, + "5": 0, + "6": 0.05 + } + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.85, + "true": 0.15 + } + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.83, + "true": 0.17 + } + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.040000000000000036, + "true": 0.96 + } + }, + { + "id": "snake:0a57c6db45c1f166083e73d8:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 1, + "south": 0, + "north": 0 + } + }, + { + "id": "snake:0a57c6db45c1f166083e73d8:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:0a57c6db45c1f166083e73d8:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:0a57c6db45c1f166083e73d8:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:63c079a7af785193d2dcb0a2:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.75, + "south": 0.23, + "east": 0.02 + } + }, + { + "id": "snake:63c079a7af785193d2dcb0a2:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "snake:63c079a7af785193d2dcb0a2:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:63c079a7af785193d2dcb0a2:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "snake:e899f3d6ecef3a09d8ffc119:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.02, + "north": 0.59, + "west": 0.39 + } + }, + { + "id": "snake:e899f3d6ecef3a09d8ffc119:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:e899f3d6ecef3a09d8ffc119:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.71, + "true": 0.29 + } + }, + { + "id": "snake:e899f3d6ecef3a09d8ffc119:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 197, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.46, + "3": 0.4, + "4": 0.03, + "5": 0.01, + "6": 0.09 + } + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.71, + "true": 0.29 + } + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.23, + "east": 0.77 + } + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 197, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.64, + "2": 0.23, + "3": 0.04, + "4": 0.01, + "5": 0, + "6": 0.08 + } + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 215, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.34, + "west": 0.07, + "east": 0.59 + } + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 196, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.25, + "3": 0.56, + "4": 0.05, + "5": 0.01, + "6": 0.13 + } + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.1, + "east": 0.9 + } + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 196, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.21, + "3": 0.62, + "4": 0.06, + "5": 0.01, + "6": 0.1 + } + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:186c9eeb1aae95c0b0c3ef44:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.72, + "south": 0.07, + "north": 0.21 + } + }, + { + "id": "snake:186c9eeb1aae95c0b0c3ef44:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "snake:186c9eeb1aae95c0b0c3ef44:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "snake:186c9eeb1aae95c0b0c3ef44:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:ce254f8faea7d887f1a666f3:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.53, + "north": 0.18, + "south": 0.29 + } + }, + { + "id": "snake:ce254f8faea7d887f1a666f3:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "snake:ce254f8faea7d887f1a666f3:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "snake:ce254f8faea7d887f1a666f3:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:09e3bf5fdec487c1c6f5af08:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 187, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.01, + "north": 0.92, + "west": 0.07 + } + }, + { + "id": "snake:09e3bf5fdec487c1c6f5af08:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:09e3bf5fdec487c1c6f5af08:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:09e3bf5fdec487c1c6f5af08:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.16, + "3": 0.68, + "4": 0.06, + "5": 0.01, + "6": 0.09 + } + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.63, + "true": 0.37 + } + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.17, + "south": 0.83 + } + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.05, + "2": 0.55, + "3": 0.26, + "4": 0.02, + "5": 0, + "6": 0.12 + } + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 230, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.29, + "north": 0.08, + "south": 0.63 + } + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.44, + "3": 0.45, + "4": 0.02, + "5": 0, + "6": 0.08 + } + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 229, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.69, + "east": 0.31 + } + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.23, + "3": 0.59, + "4": 0.07, + "5": 0.01, + "6": 0.1 + } + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:51b44ad4a37cc5094c2900be:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.45, + "south": 0.02, + "north": 0.53 + } + }, + { + "id": "snake:51b44ad4a37cc5094c2900be:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:51b44ad4a37cc5094c2900be:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:51b44ad4a37cc5094c2900be:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:b884811f4367bae3e1ebdd7a:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.6, + "north": 0.16, + "east": 0.24 + } + }, + { + "id": "snake:b884811f4367bae3e1ebdd7a:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "snake:b884811f4367bae3e1ebdd7a:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "snake:b884811f4367bae3e1ebdd7a:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "snake:dc81c3fab4d17b908531d011:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.97, + "west": 0.01, + "east": 0.02 + } + }, + { + "id": "snake:dc81c3fab4d17b908531d011:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:dc81c3fab4d17b908531d011:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:dc81c3fab4d17b908531d011:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:4a8e9c88494add4eff7bdca3:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.81, + "north": 0.03, + "east": 0.16 + } + }, + { + "id": "snake:4a8e9c88494add4eff7bdca3:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.77, + "true": 0.23 + } + }, + { + "id": "snake:4a8e9c88494add4eff7bdca3:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:4a8e9c88494add4eff7bdca3:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.76, + "2": 0.13, + "3": 0.03, + "4": 0.01, + "5": 0, + "6": 0.07 + } + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.74, + "true": 0.26 + } + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 224, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 1, + "east": 0 + } + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.78, + "2": 0.08, + "3": 0.04, + "4": 0.01, + "5": 0, + "6": 0.09 + } + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 223, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.51, + "north": 0.12, + "west": 0.37 + } + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.28, + "2": 0.35, + "3": 0.19, + "4": 0.03, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 225, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.73, + "north": 0.27 + } + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.33, + "3": 0.45999999999999996, + "4": 0.08, + "5": 0.01, + "6": 0.1 + } + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.69, + "true": 0.31 + } + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "snake:9e6867817864cb9193a8ebf1:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.38, + "north": 0.61, + "south": 0.01 + } + }, + { + "id": "snake:9e6867817864cb9193a8ebf1:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:9e6867817864cb9193a8ebf1:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:9e6867817864cb9193a8ebf1:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:22cc42ee4a862f50bec2e5b5:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.56, + "north": 0.02, + "south": 0.42 + } + }, + { + "id": "snake:22cc42ee4a862f50bec2e5b5:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.9, + "true": 0.1 + } + }, + { + "id": "snake:22cc42ee4a862f50bec2e5b5:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:22cc42ee4a862f50bec2e5b5:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:dbf416426f207e29ea1c55cf:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.62, + "west": 0.36, + "north": 0.02 + } + }, + { + "id": "snake:dbf416426f207e29ea1c55cf:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.83, + "true": 0.17 + } + }, + { + "id": "snake:dbf416426f207e29ea1c55cf:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "snake:dbf416426f207e29ea1c55cf:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:e0c7592c01c85392eb64bb73:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.99, + "north": 0.01, + "south": 0 + } + }, + { + "id": "snake:e0c7592c01c85392eb64bb73:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.87, + "true": 0.13 + } + }, + { + "id": "snake:e0c7592c01c85392eb64bb73:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:e0c7592c01c85392eb64bb73:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.1, + "3": 0.65, + "4": 0.13, + "5": 0.01, + "6": 0.11 + } + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 228, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.9, + "north": 0.07, + "west": 0.03 + } + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.17, + "2": 0.5, + "3": 0.21, + "4": 0.02, + "5": 0.01, + "6": 0.09 + } + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 227, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.07, + "west": 0.3, + "south": 0.44, + "east": 0.19 + } + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.08, + "3": 0.65, + "4": 0.18, + "5": 0.01, + "6": 0.08 + } + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 227, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.02, + "west": 0.88, + "north": 0.1 + } + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.53, + "2": 0.17, + "3": 0.09, + "4": 0.02, + "5": 0.01, + "6": 0.18 + } + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.75, + "true": 0.25 + } + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "snake:a00fe6dca724d788d40bbba7:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.01, + "east": 0.97, + "north": 0.02 + } + }, + { + "id": "snake:a00fe6dca724d788d40bbba7:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:a00fe6dca724d788d40bbba7:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:a00fe6dca724d788d40bbba7:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:04cbd162811373e57f53152b:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.63, + "west": 0.06, + "east": 0.31 + } + }, + { + "id": "snake:04cbd162811373e57f53152b:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:04cbd162811373e57f53152b:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "snake:04cbd162811373e57f53152b:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "snake:309dbe3ed47e8d1c5a17d841:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.25, + "south": 0.74, + "east": 0.01 + } + }, + { + "id": "snake:309dbe3ed47e8d1c5a17d841:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8, + "true": 0.2 + } + }, + { + "id": "snake:309dbe3ed47e8d1c5a17d841:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:309dbe3ed47e8d1c5a17d841:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "snake:5dd7b1aef1af8e9ff190595f:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.03, + "north": 0.79, + "east": 0.18 + } + }, + { + "id": "snake:5dd7b1aef1af8e9ff190595f:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:5dd7b1aef1af8e9ff190595f:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:5dd7b1aef1af8e9ff190595f:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 491, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.05, + "3": 0.38, + "4": 0.32, + "5": 0.09, + "6": 0.16 + } + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.85, + "north": 0.15 + } + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.56, + "2": 0.24, + "3": 0.06, + "4": 0.02, + "5": 0.01, + "6": 0.11 + } + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 514, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.05, + "east": 0.61, + "west": 0.34 + } + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.3, + "3": 0.39999999999999997, + "4": 0.13, + "5": 0.02, + "6": 0.14 + } + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 512, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.93, + "south": 0.07 + } + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 491, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.23, + "3": 0.29, + "4": 0.26, + "5": 0.04, + "6": 0.17 + } + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.64, + "true": 0.36 + } + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "snake:67d7d8cd30a917dd7060d049:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.13, + "south": 0.47, + "east": 0.4 + } + }, + { + "id": "snake:67d7d8cd30a917dd7060d049:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:67d7d8cd30a917dd7060d049:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:67d7d8cd30a917dd7060d049:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:850d0c734d3afba2f5d77144:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.07, + "east": 0.64, + "north": 0.29 + } + }, + { + "id": "snake:850d0c734d3afba2f5d77144:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:850d0c734d3afba2f5d77144:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "snake:850d0c734d3afba2f5d77144:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:0d69e28ac9cee01ce099a94c:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.81, + "east": 0.17, + "west": 0.02 + } + }, + { + "id": "snake:0d69e28ac9cee01ce099a94c:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:0d69e28ac9cee01ce099a94c:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:0d69e28ac9cee01ce099a94c:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:df1076fd97ec6409895096ad:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.64, + "east": 0.11, + "west": 0.25 + } + }, + { + "id": "snake:df1076fd97ec6409895096ad:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "snake:df1076fd97ec6409895096ad:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:df1076fd97ec6409895096ad:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.24, + "4": 0.48, + "5": 0.15, + "6": 0.12 + } + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6699999999999999, + "true": 0.33 + } + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.75, + "true": 0.25 + } + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.73, + "true": 0.27 + } + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.74, + "south": 0.26 + } + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.35, + "3": 0.43, + "4": 0.08, + "5": 0.01, + "6": 0.11 + } + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.75, + "true": 0.25 + } + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 510, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.05, + "east": 0.84, + "south": 0.11 + } + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.02, + "3": 0.39, + "4": 0.37, + "5": 0.11, + "6": 0.11 + } + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6599999999999999, + "true": 0.34 + } + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.96, + "south": 0.04 + } + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.77, + "2": 0.09, + "3": 0.03, + "4": 0.01, + "5": 0, + "6": 0.09999999999999999 + } + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.63, + "true": 0.37 + } + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:09d01d5cb2826ee7d7b75692:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.29, + "south": 0.07, + "east": 0.64 + } + }, + { + "id": "snake:09d01d5cb2826ee7d7b75692:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:09d01d5cb2826ee7d7b75692:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:09d01d5cb2826ee7d7b75692:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:6522cd24159d9f67a110fb91:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9999999999999999, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.57, + "north": 0.41, + "east": 0.02 + } + }, + { + "id": "snake:6522cd24159d9f67a110fb91:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:6522cd24159d9f67a110fb91:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "snake:6522cd24159d9f67a110fb91:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "snake:8a431ce79c63ab176847b014:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.88, + "west": 0.03, + "east": 0.09 + } + }, + { + "id": "snake:8a431ce79c63ab176847b014:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:8a431ce79c63ab176847b014:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "snake:8a431ce79c63ab176847b014:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8, + "true": 0.2 + } + }, + { + "id": "snake:5fef7acaca5a9f9ea2b6dce3:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.01, + "south": 0.23, + "east": 0.76 + } + }, + { + "id": "snake:5fef7acaca5a9f9ea2b6dce3:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.86, + "true": 0.14 + } + }, + { + "id": "snake:5fef7acaca5a9f9ea2b6dce3:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:5fef7acaca5a9f9ea2b6dce3:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 484, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.3, + "4": 0.46, + "5": 0.12, + "6": 0.11 + } + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 518, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 518, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 518, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.7, + "true": 0.3 + } + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 518, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.53, + "west": 0.47 + } + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.42, + "3": 0.4, + "4": 0.07, + "5": 0.01, + "6": 0.08 + } + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 506, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.17, + "south": 0.67, + "east": 0.16 + } + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.4, + "4": 0.43, + "5": 0.08, + "6": 0.08 + } + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 505, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.02, + "north": 0.09, + "east": 0.89 + } + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 491, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.05, + "3": 0.46, + "4": 0.36, + "5": 0.05, + "6": 0.08 + } + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6699999999999999, + "true": 0.33 + } + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:92c385cd4eb8c43cda20267f:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.24, + "north": 0.13, + "east": 0.63 + } + }, + { + "id": "snake:92c385cd4eb8c43cda20267f:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:92c385cd4eb8c43cda20267f:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:92c385cd4eb8c43cda20267f:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:1c07f3a9c4269d3d224fa63b:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.02, + "south": 0.01, + "west": 0.97 + } + }, + { + "id": "snake:1c07f3a9c4269d3d224fa63b:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:1c07f3a9c4269d3d224fa63b:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "snake:1c07f3a9c4269d3d224fa63b:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:6bac4198ff50afdd1c28bf82:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 186, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.25, + "west": 0.52, + "north": 0.23 + } + }, + { + "id": "snake:6bac4198ff50afdd1c28bf82:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:6bac4198ff50afdd1c28bf82:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "snake:6bac4198ff50afdd1c28bf82:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:4d5d9a736a2c010ecddf3bd7:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.11, + "west": 0.02, + "north": 0.87 + } + }, + { + "id": "snake:4d5d9a736a2c010ecddf3bd7:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:4d5d9a736a2c010ecddf3bd7:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:4d5d9a736a2c010ecddf3bd7:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 505, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.3, + "4": 0.33999999999999997, + "5": 0.11, + "6": 0.24 + } + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.88, + "true": 0.12 + } + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.84, + "true": 0.16 + } + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 520, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.06, + "east": 0.93, + "south": 0.01 + } + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 500, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 506, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.59, + "2": 0.27, + "3": 0.05, + "4": 0.01, + "5": 0, + "6": 0.08 + } + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 519, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.86, + "west": 0.02, + "south": 0.12 + } + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 505, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.03, + "3": 0.62, + "4": 0.23, + "5": 0.02, + "6": 0.1 + } + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 518, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.06, + "north": 0.27, + "east": 0.67 + } + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 505, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.45, + "4": 0.39, + "5": 0.05, + "6": 0.1 + } + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "snake:ec8cd8b3711ab660e4121bfd:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.13, + "east": 0.72, + "north": 0.15 + } + }, + { + "id": "snake:ec8cd8b3711ab660e4121bfd:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:ec8cd8b3711ab660e4121bfd:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:ec8cd8b3711ab660e4121bfd:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:611b5d26ce10b988029a6daf:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0, + "east": 1, + "north": 0 + } + }, + { + "id": "snake:611b5d26ce10b988029a6daf:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:611b5d26ce10b988029a6daf:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:611b5d26ce10b988029a6daf:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "snake:a408545a8ce348a3053aa8ed:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.74, + "south": 0.16, + "north": 0.1 + } + }, + { + "id": "snake:a408545a8ce348a3053aa8ed:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "snake:a408545a8ce348a3053aa8ed:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:a408545a8ce348a3053aa8ed:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:9828eefaea5a58f642317f6e:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.01, + "west": 0.97, + "north": 0.02 + } + }, + { + "id": "snake:9828eefaea5a58f642317f6e:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "snake:9828eefaea5a58f642317f6e:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "snake:9828eefaea5a58f642317f6e:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "snake:0767d5e59aed5fd05cac2e9f:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.15, + "east": 0.62, + "north": 0.23 + } + }, + { + "id": "snake:0767d5e59aed5fd05cac2e9f:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:0767d5e59aed5fd05cac2e9f:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:0767d5e59aed5fd05cac2e9f:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:4f81c7dda2b7d32567d91775:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.13, + "north": 0.08, + "east": 0.79 + } + }, + { + "id": "snake:4f81c7dda2b7d32567d91775:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "snake:4f81c7dda2b7d32567d91775:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:4f81c7dda2b7d32567d91775:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "snake:852918c7ae73c068c0ff56c9:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.16, + "north": 0.13, + "west": 0.71 + } + }, + { + "id": "snake:852918c7ae73c068c0ff56c9:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:852918c7ae73c068c0ff56c9:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "snake:852918c7ae73c068c0ff56c9:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:635d2abfc307ec0710e44f69:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.98, + "east": 0.02, + "west": 0 + } + }, + { + "id": "snake:635d2abfc307ec0710e44f69:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:635d2abfc307ec0710e44f69:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:635d2abfc307ec0710e44f69:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 128, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.06, + "2": 0.61, + "3": 0.1, + "4": 0.01, + "5": 0.01, + "6": 0.21 + } + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.74, + "true": 0.26 + } + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 148, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 1, + "east": 0 + } + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 129, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.8, + "2": 0.1, + "3": 0.02, + "4": 0, + "5": 0, + "6": 0.08 + } + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.74, + "true": 0.26 + } + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 147, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.17, + "west": 0.83 + } + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 128, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.17, + "2": 0.54, + "3": 0.05, + "4": 0.01, + "5": 0, + "6": 0.23 + } + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.74, + "true": 0.26 + } + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.7, + "true": 0.3 + } + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 147, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.15, + "east": 0.85 + } + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 128, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.64, + "2": 0.17, + "3": 0.02, + "4": 0, + "5": 0, + "6": 0.17 + } + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.65, + "true": 0.35 + } + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.63, + "true": 0.37 + } + }, + { + "id": "snake:6b804b8015033d9a92a4dec1:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 1, + "north": 0, + "south": 0 + } + }, + { + "id": "snake:6b804b8015033d9a92a4dec1:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:6b804b8015033d9a92a4dec1:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:6b804b8015033d9a92a4dec1:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:c36d051be7af4b360d34b50d:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.01, + "north": 0.45, + "west": 0.54 + } + }, + { + "id": "snake:c36d051be7af4b360d34b50d:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:c36d051be7af4b360d34b50d:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "snake:c36d051be7af4b360d34b50d:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "snake:8cac327e96392079d7260378:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.61, + "west": 0.29, + "east": 0.1 + } + }, + { + "id": "snake:8cac327e96392079d7260378:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:8cac327e96392079d7260378:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:8cac327e96392079d7260378:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:0d2c2cfdf6e63083ab122e25:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.05, + "south": 0.89, + "east": 0.06 + } + }, + { + "id": "snake:0d2c2cfdf6e63083ab122e25:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "snake:0d2c2cfdf6e63083ab122e25:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "snake:0d2c2cfdf6e63083ab122e25:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.46, + "2": 0.25, + "3": 0.05, + "4": 0.01, + "5": 0.01, + "6": 0.22 + } + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.69, + "true": 0.31 + } + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.83, + "true": 0.17 + } + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.64, + "true": 0.36 + } + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.97, + "north": 0.03 + } + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.65, + "2": 0.19, + "3": 0.02, + "4": 0, + "5": 0, + "6": 0.14 + } + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.79, + "true": 0.21 + } + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 156, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.64, + "west": 0.05, + "south": 0.31 + } + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 143, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.09, + "2": 0.57, + "3": 0.07, + "4": 0.01, + "5": 0.01, + "6": 0.25 + } + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.04, + "south": 0.96 + } + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.4, + "2": 0.41, + "3": 0.03, + "4": 0.01, + "5": 0, + "6": 0.15 + } + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:de8b76d31b3502f06ab7bece:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.19, + "east": 0.59, + "north": 0.22 + } + }, + { + "id": "snake:de8b76d31b3502f06ab7bece:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:de8b76d31b3502f06ab7bece:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:de8b76d31b3502f06ab7bece:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:b6f0bcae75ff0eedc4104dd0:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.01, + "west": 0.64, + "south": 0.35 + } + }, + { + "id": "snake:b6f0bcae75ff0eedc4104dd0:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.77, + "true": 0.23 + } + }, + { + "id": "snake:b6f0bcae75ff0eedc4104dd0:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "snake:b6f0bcae75ff0eedc4104dd0:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "snake:2340031ad35a3132a389f44d:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.02, + "south": 0.28, + "east": 0.7 + } + }, + { + "id": "snake:2340031ad35a3132a389f44d:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.89, + "true": 0.11 + } + }, + { + "id": "snake:2340031ad35a3132a389f44d:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:2340031ad35a3132a389f44d:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:d76b23c4185e25ab14dda3c8:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.09, + "east": 0.77, + "south": 0.14 + } + }, + { + "id": "snake:d76b23c4185e25ab14dda3c8:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.75, + "true": 0.25 + } + }, + { + "id": "snake:d76b23c4185e25ab14dda3c8:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "snake:d76b23c4185e25ab14dda3c8:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.54, + "3": 0.37, + "4": 0.02, + "5": 0, + "6": 0.07 + } + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.87, + "true": 0.13 + } + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.27, + "south": 0.7, + "east": 0.03 + } + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.52, + "2": 0.34, + "3": 0.04, + "4": 0.01, + "5": 0, + "6": 0.09 + } + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.48, + "south": 0.49, + "north": 0.03 + } + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.51, + "3": 0.4, + "4": 0.02, + "5": 0, + "6": 0.07 + } + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.88, + "true": 0.12 + } + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.51, + "west": 0.42, + "east": 0.02, + "north": 0.05 + } + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.39, + "2": 0.49, + "3": 0.04, + "4": 0, + "5": 0, + "6": 0.08 + } + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:a9a97a95a878d441fd0eed1d:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0, + "north": 0.26, + "east": 0.74 + } + }, + { + "id": "snake:a9a97a95a878d441fd0eed1d:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:a9a97a95a878d441fd0eed1d:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:a9a97a95a878d441fd0eed1d:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:f4916f37b65f9a17a6e6d10b:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9999999999999999, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.01, + "west": 0.69, + "north": 0.3 + } + }, + { + "id": "snake:f4916f37b65f9a17a6e6d10b:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.63, + "true": 0.37 + } + }, + { + "id": "snake:f4916f37b65f9a17a6e6d10b:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:f4916f37b65f9a17a6e6d10b:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:1f964939ebb08f4325841e5b:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.54, + "west": 0.45, + "south": 0.01 + } + }, + { + "id": "snake:1f964939ebb08f4325841e5b:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "snake:1f964939ebb08f4325841e5b:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.77, + "true": 0.23 + } + }, + { + "id": "snake:1f964939ebb08f4325841e5b:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:58485cd9064d09b8277665a0:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.99, + "west": 0.01, + "east": 0 + } + }, + { + "id": "snake:58485cd9064d09b8277665a0:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.83, + "true": 0.17 + } + }, + { + "id": "snake:58485cd9064d09b8277665a0:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "snake:58485cd9064d09b8277665a0:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 226, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.27, + "3": 0.48, + "4": 0.08, + "5": 0.01, + "6": 0.15 + } + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.28, + "east": 0.72 + } + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 226, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.38, + "3": 0.39999999999999997, + "4": 0.08, + "5": 0.01, + "6": 0.12 + } + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 240, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.06, + "south": 0.25, + "west": 0.69 + } + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 226, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.49, + "2": 0.31, + "3": 0.06, + "4": 0.01, + "5": 0.01, + "6": 0.12 + } + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.59, + "east": 0.41 + } + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 226, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.57, + "2": 0.14, + "3": 0.08, + "4": 0.02, + "5": 0.01, + "6": 0.18 + } + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:e71d44addec992f5c8c8e8f0:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0, + "north": 0.99, + "east": 0.01 + } + }, + { + "id": "snake:e71d44addec992f5c8c8e8f0:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:e71d44addec992f5c8c8e8f0:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:e71d44addec992f5c8c8e8f0:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:d50907af6656e0c98fbf31da:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.01, + "east": 0.74, + "south": 0.25 + } + }, + { + "id": "snake:d50907af6656e0c98fbf31da:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "snake:d50907af6656e0c98fbf31da:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:d50907af6656e0c98fbf31da:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "snake:e97eaa71c50f430fe64471d5:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.99, + "north": 0.01, + "south": 0 + } + }, + { + "id": "snake:e97eaa71c50f430fe64471d5:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "snake:e97eaa71c50f430fe64471d5:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "snake:e97eaa71c50f430fe64471d5:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.14, + "3": 0.61, + "4": 0.12, + "5": 0.01, + "6": 0.12 + } + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.12, + "south": 0.88 + } + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.03, + "2": 0.47, + "3": 0.28, + "4": 0.04, + "5": 0.01, + "6": 0.17 + } + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.61, + "true": 0.39 + } + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.8200000000000001, + "true": 0.18 + } + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 229, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.25, + "north": 0.04, + "east": 0.71 + } + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.33, + "3": 0.5, + "4": 0.05, + "5": 0.01, + "6": 0.1 + } + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.11, + "west": 0.89 + } + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 219, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.25, + "3": 0.49, + "4": 0.08, + "5": 0.01, + "6": 0.16 + } + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 247, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 247, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 247, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 247, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:f2f72dad5dedbdb36adc45a1:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.5, + "north": 0.48, + "south": 0.02 + } + }, + { + "id": "snake:f2f72dad5dedbdb36adc45a1:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:f2f72dad5dedbdb36adc45a1:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:f2f72dad5dedbdb36adc45a1:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:4908b0108026b8d03e2e30b2:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.04, + "south": 0.1, + "east": 0.86 + } + }, + { + "id": "snake:4908b0108026b8d03e2e30b2:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "snake:4908b0108026b8d03e2e30b2:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "snake:4908b0108026b8d03e2e30b2:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "snake:9e21f7a6a1b5a3f6776d566b:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.25, + "east": 0.02, + "west": 0.73 + } + }, + { + "id": "snake:9e21f7a6a1b5a3f6776d566b:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:9e21f7a6a1b5a3f6776d566b:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:9e21f7a6a1b5a3f6776d566b:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:c60247f3751938fe95596441:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.67, + "east": 0.02, + "north": 0.31 + } + }, + { + "id": "snake:c60247f3751938fe95596441:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "snake:c60247f3751938fe95596441:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "snake:c60247f3751938fe95596441:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.13, + "east": 0.87 + } + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 220, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.04, + "3": 0.71, + "4": 0.1, + "5": 0.01, + "6": 0.14 + } + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 235, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.17, + "south": 0.77, + "north": 0.06 + } + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.13, + "2": 0.54, + "3": 0.16, + "4": 0.02, + "5": 0, + "6": 0.15 + } + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.63, + "true": 0.37 + } + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 234, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.08, + "east": 0.49, + "south": 0.43 + } + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.04, + "3": 0.7, + "4": 0.1, + "5": 0.01, + "6": 0.15 + } + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.96, + "north": 0.04 + } + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 220, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.04, + "2": 0.37, + "3": 0.32, + "4": 0.07, + "5": 0.01, + "6": 0.19 + } + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.76, + "true": 0.24 + } + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "snake:e272f5b22e1649dc9a40f56f:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.74, + "north": 0.01, + "south": 0.25 + } + }, + { + "id": "snake:e272f5b22e1649dc9a40f56f:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:e272f5b22e1649dc9a40f56f:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "snake:e272f5b22e1649dc9a40f56f:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:1a75df6d584a4d37af00c4f9:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.82, + "south": 0.17, + "east": 0.01 + } + }, + { + "id": "snake:1a75df6d584a4d37af00c4f9:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.83, + "true": 0.17 + } + }, + { + "id": "snake:1a75df6d584a4d37af00c4f9:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:1a75df6d584a4d37af00c4f9:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "snake:ba01f7e9755f2a8bc3f51ad5:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.02, + "north": 0.38, + "east": 0.6 + } + }, + { + "id": "snake:ba01f7e9755f2a8bc3f51ad5:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:ba01f7e9755f2a8bc3f51ad5:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "snake:ba01f7e9755f2a8bc3f51ad5:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.83, + "true": 0.17 + } + }, + { + "id": "snake:ad228b61d8212dc6c4cffdf1:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.04, + "east": 0.1, + "south": 0.86 + } + }, + { + "id": "snake:ad228b61d8212dc6c4cffdf1:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "snake:ad228b61d8212dc6c4cffdf1:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "snake:ad228b61d8212dc6c4cffdf1:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.81, + "true": 0.19 + } + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.35, + "3": 0.5, + "4": 0.05, + "5": 0.01, + "6": 0.07 + } + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 237, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.01, + "north": 0.47, + "east": 0.52, + "south": 0 + } + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0.01, + "1": 0.79, + "2": 0.07, + "3": 0.02, + "4": 0.01, + "5": 0, + "6": 0.1 + } + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 236, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.3, + "west": 0.12, + "north": 0.05, + "south": 0.53 + } + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.44, + "3": 0.44, + "4": 0.05, + "5": 0, + "6": 0.06 + } + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.8, + "south": 0.2 + } + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.2, + "3": 0.67, + "4": 0.06, + "5": 0, + "6": 0.07 + } + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:af8239c080fc2cb3fba98d2b:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.1, + "south": 0.14, + "east": 0.76 + } + }, + { + "id": "snake:af8239c080fc2cb3fba98d2b:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:af8239c080fc2cb3fba98d2b:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:af8239c080fc2cb3fba98d2b:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.17000000000000004, + "true": 0.83 + } + }, + { + "id": "snake:9c812cad9f595a262b7d3e62:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0, + "south": 0.99, + "east": 0.01 + } + }, + { + "id": "snake:9c812cad9f595a262b7d3e62:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:9c812cad9f595a262b7d3e62:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:9c812cad9f595a262b7d3e62:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "snake:243bdce11167f2cb54cd23a6:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.82, + "east": 0.15, + "north": 0.03 + } + }, + { + "id": "snake:243bdce11167f2cb54cd23a6:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:243bdce11167f2cb54cd23a6:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:243bdce11167f2cb54cd23a6:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "snake:e2c971a95e7b505a5e37f7d9:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.99, + "east": 0.01, + "west": 0 + } + }, + { + "id": "snake:e2c971a95e7b505a5e37f7d9:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:e2c971a95e7b505a5e37f7d9:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:e2c971a95e7b505a5e37f7d9:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 474, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 480, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.06, + "3": 0.25, + "4": 0.44, + "5": 0.09, + "6": 0.16 + } + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.07, + "south": 0.93 + } + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 473, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 479, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.17, + "2": 0.41, + "3": 0.24, + "4": 0.04, + "5": 0.01, + "6": 0.13 + } + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6799999999999999, + "true": 0.32 + } + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 495, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.85, + "east": 0.11, + "south": 0.04 + } + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 474, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 480, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.09, + "3": 0.36, + "4": 0.34, + "5": 0.05, + "6": 0.16 + } + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.7, + "true": 0.3 + } + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.25, + "south": 0.75 + } + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 473, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 479, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.21, + "3": 0.44, + "4": 0.21, + "5": 0.02, + "6": 0.12 + } + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "snake:e194fc071b1fde4b9e783c83:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.43, + "north": 0.53, + "south": 0.04 + } + }, + { + "id": "snake:e194fc071b1fde4b9e783c83:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:e194fc071b1fde4b9e783c83:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:e194fc071b1fde4b9e783c83:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:fcf919116a7028c2c3286bf4:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.97, + "north": 0.02, + "south": 0.01 + } + }, + { + "id": "snake:fcf919116a7028c2c3286bf4:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:fcf919116a7028c2c3286bf4:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "snake:fcf919116a7028c2c3286bf4:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:99b3c3c8700d99220a1d030c:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 185, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.64, + "north": 0.04, + "south": 0.32 + } + }, + { + "id": "snake:99b3c3c8700d99220a1d030c:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:99b3c3c8700d99220a1d030c:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:99b3c3c8700d99220a1d030c:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:17afc77a551ef63112fd0716:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.06, + "south": 0.01, + "east": 0.93 + } + }, + { + "id": "snake:17afc77a551ef63112fd0716:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "snake:17afc77a551ef63112fd0716:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:17afc77a551ef63112fd0716:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 518, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.04, + "3": 0.37, + "4": 0.35, + "5": 0.09, + "6": 0.15 + } + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 552, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 552, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 552, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 552, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.08, + "south": 0.92 + } + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 517, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.24, + "true": 0.76 + } + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 523, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.06, + "3": 0.44, + "4": 0.35, + "5": 0.05, + "6": 0.1 + } + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 551, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.64, + "true": 0.36 + } + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 551, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 551, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.74, + "true": 0.26 + } + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 551, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.52, + "true": 0.48 + } + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 539, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.34, + "east": 0.39, + "north": 0.27 + } + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.47000000000000003, + "3": 0.31, + "4": 0.07, + "5": 0.01, + "6": 0.13 + } + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.44999999999999996, + "true": 0.55 + } + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 540, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.84, + "north": 0.16 + } + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.19, + "3": 0.46, + "4": 0.21, + "5": 0.02, + "6": 0.12 + } + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.62, + "true": 0.38 + } + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.78, + "true": 0.22 + } + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "snake:4a0aee04affa084fcaeeaf52:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.12, + "south": 0.4, + "east": 0.48 + } + }, + { + "id": "snake:4a0aee04affa084fcaeeaf52:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:4a0aee04affa084fcaeeaf52:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:4a0aee04affa084fcaeeaf52:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:3b5df35292fcfcfb1d2eeb27:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.99, + "west": 0, + "east": 0.01 + } + }, + { + "id": "snake:3b5df35292fcfcfb1d2eeb27:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:3b5df35292fcfcfb1d2eeb27:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:3b5df35292fcfcfb1d2eeb27:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "snake:6a6399d8e17bf0a9bc5100f2:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.63, + "north": 0.1, + "south": 0.27 + } + }, + { + "id": "snake:6a6399d8e17bf0a9bc5100f2:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "snake:6a6399d8e17bf0a9bc5100f2:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "snake:6a6399d8e17bf0a9bc5100f2:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "snake:3dea7e463701bc106a5b9e80:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.04, + "south": 0.7, + "west": 0.26 + } + }, + { + "id": "snake:3dea7e463701bc106a5b9e80:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:3dea7e463701bc106a5b9e80:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:3dea7e463701bc106a5b9e80:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 513, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.22, + "south": 0.78 + } + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.22, + "4": 0.44, + "5": 0.16, + "6": 0.17 + } + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.65, + "true": 0.35 + } + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.77, + "true": 0.23 + } + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 513, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.95, + "north": 0.05 + } + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.43, + "3": 0.24000000000000002, + "4": 0.13, + "5": 0.03, + "6": 0.15 + } + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.79, + "true": 0.21 + } + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 515, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.16, + "west": 0.68, + "north": 0.16 + } + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 500, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.03, + "3": 0.25, + "4": 0.41000000000000003, + "5": 0.14, + "6": 0.17 + } + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.53, + "true": 0.47 + } + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.74, + "true": 0.26 + } + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.22, + "south": 0.78 + } + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 491, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.2, + "4": 0.47000000000000003, + "5": 0.17, + "6": 0.15 + } + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.63, + "true": 0.37 + } + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.77, + "true": 0.23 + } + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5700000000000001, + "true": 0.43 + } + }, + { + "id": "snake:9f940aae7bd55ce7e711ac49:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.03, + "north": 0.4, + "east": 0.57 + } + }, + { + "id": "snake:9f940aae7bd55ce7e711ac49:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:9f940aae7bd55ce7e711ac49:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:9f940aae7bd55ce7e711ac49:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "snake:53c433339e8c5e57ea898a50:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 177, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.4, + "west": 0.01, + "east": 0.59 + } + }, + { + "id": "snake:53c433339e8c5e57ea898a50:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.54, + "true": 0.46 + } + }, + { + "id": "snake:53c433339e8c5e57ea898a50:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:53c433339e8c5e57ea898a50:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "snake:9e1043e67de35a1943feb328:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.64, + "east": 0.04, + "west": 0.32 + } + }, + { + "id": "snake:9e1043e67de35a1943feb328:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:9e1043e67de35a1943feb328:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:9e1043e67de35a1943feb328:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:780daa690f231eb74dd97dca:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.18, + "north": 0.3, + "west": 0.52 + } + }, + { + "id": "snake:780daa690f231eb74dd97dca:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "snake:780daa690f231eb74dd97dca:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "snake:780daa690f231eb74dd97dca:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 459, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 465, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0, + "3": 0.38, + "4": 0.39, + "5": 0.11, + "6": 0.12 + } + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.74, + "true": 0.26 + } + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.9, + "true": 0.1 + } + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 480, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.47, + "east": 0.02, + "north": 0.05, + "west": 0.46 + } + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 459, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 465, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.24, + "3": 0.54, + "4": 0.08, + "5": 0.02, + "6": 0.1 + } + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.19999999999999996, + "true": 0.8 + } + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.14, + "true": 0.86 + } + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 480, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.06, + "south": 0.3, + "west": 0.64 + } + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 460, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 466, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0, + "3": 0.33, + "4": 0.45, + "5": 0.11, + "6": 0.11 + } + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.49, + "true": 0.51 + } + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.89, + "true": 0.11 + } + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 481, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.06, + "south": 0.49, + "west": 0.45 + } + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 460, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 466, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.1, + "3": 0.47000000000000003, + "4": 0.27, + "5": 0.06, + "6": 0.1 + } + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.81, + "true": 0.19 + } + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:7a64e76f56a37f3684f5f79f:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.54, + "south": 0.3, + "north": 0.16 + } + }, + { + "id": "snake:7a64e76f56a37f3684f5f79f:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:7a64e76f56a37f3684f5f79f:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:7a64e76f56a37f3684f5f79f:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:70a97174b2c166c712894d0b:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0, + "east": 0.01, + "south": 0.99 + } + }, + { + "id": "snake:70a97174b2c166c712894d0b:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:70a97174b2c166c712894d0b:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:70a97174b2c166c712894d0b:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "snake:57971912a78a6cac83176839:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 187, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.15, + "north": 0.84, + "west": 0.01 + } + }, + { + "id": "snake:57971912a78a6cac83176839:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:57971912a78a6cac83176839:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:57971912a78a6cac83176839:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:fd0cb58bb748915f35c51f85:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 1, + "south": 0, + "north": 0 + } + }, + { + "id": "snake:fd0cb58bb748915f35c51f85:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:fd0cb58bb748915f35c51f85:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "snake:fd0cb58bb748915f35c51f85:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 730, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 736, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.03, + "3": 0.31, + "4": 0.3, + "5": 0.14, + "6": 0.22 + } + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.65, + "true": 0.35 + } + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 751, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.87, + "west": 0.13 + } + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 730, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 736, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.09, + "2": 0.37, + "3": 0.25, + "4": 0.09, + "5": 0.04, + "6": 0.16 + } + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5900000000000001, + "true": 0.41 + } + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.22999999999999998, + "true": 0.77 + } + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 751, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.45, + "north": 0.13, + "south": 0.42 + } + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 730, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 736, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.02, + "3": 0.3, + "4": 0.28, + "5": 0.14, + "6": 0.26 + } + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43000000000000005, + "true": 0.57 + } + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.7, + "true": 0.3 + } + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.64, + "true": 0.36 + } + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 752, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.86, + "east": 0.14 + } + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 731, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 737, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.41, + "3": 0.23, + "4": 0.13, + "5": 0.04, + "6": 0.18 + } + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 765, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.47, + "true": 0.53 + } + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 765, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 765, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 765, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:3457a5bcfc7ae7e65820c327:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.02, + "east": 0.5700000000000001, + "north": 0.41 + } + }, + { + "id": "snake:3457a5bcfc7ae7e65820c327:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.10999999999999999, + "true": 0.89 + } + }, + { + "id": "snake:3457a5bcfc7ae7e65820c327:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:3457a5bcfc7ae7e65820c327:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:f2b50f863c17077d11d4476a:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.01, + "north": 0.04, + "west": 0.95 + } + }, + { + "id": "snake:f2b50f863c17077d11d4476a:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:f2b50f863c17077d11d4476a:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "snake:f2b50f863c17077d11d4476a:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:89685da0ba3d37f15f6a8d29:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.02, + "east": 0.11, + "north": 0.87 + } + }, + { + "id": "snake:89685da0ba3d37f15f6a8d29:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:89685da0ba3d37f15f6a8d29:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "snake:89685da0ba3d37f15f6a8d29:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "snake:a2aea55c8183660f23126df4:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 185, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.06, + "east": 0.26, + "south": 0.6799999999999999 + } + }, + { + "id": "snake:a2aea55c8183660f23126df4:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "snake:a2aea55c8183660f23126df4:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:a2aea55c8183660f23126df4:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5800000000000001, + "true": 0.42 + } + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1063, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1069, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.21, + "4": 0.4, + "5": 0.24, + "6": 0.14 + } + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.5, + "true": 0.5 + } + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 1085, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.5, + "south": 0.5 + } + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1064, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1070, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.02, + "2": 0.41000000000000003, + "3": 0.33, + "4": 0.11, + "5": 0.02, + "6": 0.11 + } + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 1082, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.18, + "east": 0.66, + "north": 0.16 + } + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1063, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1069, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.16, + "4": 0.46, + "5": 0.24, + "6": 0.13 + } + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 1085, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.58, + "north": 0.42 + } + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1064, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1070, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.09, + "3": 0.36, + "4": 0.33, + "5": 0.08, + "6": 0.14 + } + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6699999999999999, + "true": 0.33 + } + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:4323c20b90266bb4d9d253f9:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.03, + "north": 0.47, + "east": 0.5 + } + }, + { + "id": "snake:4323c20b90266bb4d9d253f9:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:4323c20b90266bb4d9d253f9:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:4323c20b90266bb4d9d253f9:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:06405ef35d5208c85b1ebb1d:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.19, + "south": 0.03, + "west": 0.78 + } + }, + { + "id": "snake:06405ef35d5208c85b1ebb1d:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "snake:06405ef35d5208c85b1ebb1d:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "snake:06405ef35d5208c85b1ebb1d:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:a17f85e8992b92bfc18067aa:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 185, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.01, + "south": 0.4, + "east": 0.59 + } + }, + { + "id": "snake:a17f85e8992b92bfc18067aa:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "snake:a17f85e8992b92bfc18067aa:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.20999999999999996, + "true": 0.79 + } + }, + { + "id": "snake:a17f85e8992b92bfc18067aa:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "snake:d74471f2cd14288707392991:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.2, + "west": 0.79, + "east": 0.01 + } + }, + { + "id": "snake:d74471f2cd14288707392991:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.29000000000000004, + "true": 0.71 + } + }, + { + "id": "snake:d74471f2cd14288707392991:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:d74471f2cd14288707392991:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.41000000000000003, + "true": 0.59 + } + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1053, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.39, + "true": 0.61 + } + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1059, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0.01, + "2": 0.01, + "3": 0.09, + "4": 0.21, + "5": 0.36, + "6": 0.32 + } + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.42000000000000004, + "true": 0.58 + } + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.45999999999999996, + "true": 0.54 + } + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.78, + "true": 0.22 + } + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 1073, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.01, + "west": 0.95, + "north": 0.04 + } + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1053, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1059, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 0.9900000000000001, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": "Rounded target has no identity probability-simplex representative", + "teacher_raw_mapping": { + "0": 0, + "1": 0.07, + "2": 0.56, + "3": 0.17, + "4": 0.03, + "5": 0.02, + "6": 0.14 + } + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.56, + "true": 0.44 + } + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 1074, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.4, + "north": 0.24, + "east": 0.36 + } + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1054, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1060, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.23, + "4": 0.28, + "5": 0.27, + "6": 0.21 + } + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.55, + "true": 0.45 + } + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 1075, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.22, + "north": 0.2, + "west": 0.58 + } + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1054, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1060, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.2, + "4": 0.35, + "5": 0.27, + "6": 0.17 + } + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.36, + "true": 0.64 + } + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.27, + "true": 0.73 + } + }, + { + "id": "snake:48047719c35c8036f2597b4c:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.15, + "south": 0.07, + "east": 0.78 + } + }, + { + "id": "snake:48047719c35c8036f2597b4c:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:48047719c35c8036f2597b4c:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:48047719c35c8036f2597b4c:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.16000000000000003, + "true": 0.84 + } + }, + { + "id": "snake:6b7a640b5bdf1f583cf06e15:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.6, + "west": 0.34, + "east": 0.06 + } + }, + { + "id": "snake:6b7a640b5bdf1f583cf06e15:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "snake:6b7a640b5bdf1f583cf06e15:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.07999999999999996, + "true": 0.92 + } + }, + { + "id": "snake:6b7a640b5bdf1f583cf06e15:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.48, + "true": 0.52 + } + }, + { + "id": "snake:20906b8454960dd9e5ef74c3:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.29, + "north": 0.67, + "east": 0.04 + } + }, + { + "id": "snake:20906b8454960dd9e5ef74c3:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.12, + "true": 0.88 + } + }, + { + "id": "snake:20906b8454960dd9e5ef74c3:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "snake:20906b8454960dd9e5ef74c3:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "snake:a2c757cd44b5e0873f48c513:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.7, + "east": 0.03, + "south": 0.27 + } + }, + { + "id": "snake:a2c757cd44b5e0873f48c513:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + }, + { + "id": "snake:a2c757cd44b5e0873f48c513:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "snake:a2c757cd44b5e0873f48c513:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.28, + "true": 0.72 + } + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 1027, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "west": 0.93, + "east": 0.07 + } + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1007, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1013, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0, + "3": 0.07, + "4": 0.39, + "5": 0.37, + "6": 0.17 + } + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.37, + "true": 0.63 + } + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.6, + "true": 0.4 + } + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 1026, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.13, + "west": 0.71, + "north": 0.16 + } + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1006, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.30000000000000004, + "true": 0.7 + } + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1012, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0.01, + "1": 0.73, + "2": 0.04, + "3": 0.03, + "4": 0.01, + "5": 0.01, + "6": 0.17 + } + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31000000000000005, + "true": 0.69 + } + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.35, + "true": 0.65 + } + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 1026, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.65, + "west": 0.3, + "east": 0.05 + } + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1006, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1012, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0, + "3": 0.17, + "4": 0.41000000000000003, + "5": 0.29, + "6": 0.13 + } + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.26, + "true": 0.74 + } + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.33999999999999997, + "true": 0.66 + } + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18999999999999995, + "true": 0.81 + } + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 1028, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.17, + "west": 0.83 + } + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1007, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.32999999999999996, + "true": 0.67 + } + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1013, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "0": 0, + "1": 0, + "2": 0.01, + "3": 0.39, + "4": 0.35, + "5": 0.13, + "6": 0.12 + } + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.38, + "true": 0.62 + } + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.4, + "true": 0.6 + } + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.79, + "true": 0.21 + } + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.31999999999999995, + "true": 0.68 + } + }, + { + "id": "snake:7928813fe2cda7456d1a2d24:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0.09, + "north": 0.24, + "east": 0.67 + } + }, + { + "id": "snake:7928813fe2cda7456d1a2d24:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.09999999999999998, + "true": 0.9 + } + }, + { + "id": "snake:7928813fe2cda7456d1a2d24:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06000000000000005, + "true": 0.94 + } + }, + { + "id": "snake:7928813fe2cda7456d1a2d24:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.13, + "true": 0.87 + } + }, + { + "id": "snake:56fae55f94b19332a280e632:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "south": 0, + "north": 0, + "west": 1 + } + }, + { + "id": "snake:56fae55f94b19332a280e632:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.06999999999999995, + "true": 0.93 + } + }, + { + "id": "snake:56fae55f94b19332a280e632:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.15000000000000002, + "true": 0.85 + } + }, + { + "id": "snake:56fae55f94b19332a280e632:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.050000000000000044, + "true": 0.95 + } + }, + { + "id": "snake:910b45e72218ee95ef8b49f2:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "north": 0.7, + "east": 0.05, + "west": 0.25 + } + }, + { + "id": "snake:910b45e72218ee95ef8b49f2:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.18000000000000005, + "true": 0.82 + } + }, + { + "id": "snake:910b45e72218ee95ef8b49f2:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": false, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.51, + "true": 0.49 + } + }, + { + "id": "snake:910b45e72218ee95ef8b49f2:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.25, + "true": 0.75 + } + }, + { + "id": "snake:fa0d0b06be9f8698a9622a2b:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "east": 0.89, + "south": 0.1, + "north": 0.01 + } + }, + { + "id": "snake:fa0d0b06be9f8698a9622a2b:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.21999999999999997, + "true": 0.78 + } + }, + { + "id": "snake:fa0d0b06be9f8698a9622a2b:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.08999999999999997, + "true": 0.91 + } + }, + { + "id": "snake:fa0d0b06be9f8698a9622a2b:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": 1.0, + "teacher_proxy_usable": true, + "teacher_argmax_matches_gold": true, + "target_transform": "identity_rounded_proxy", + "teacher_target_error": null, + "teacher_raw_mapping": { + "false": 0.43999999999999995, + "true": 0.56 + } + } +] diff --git a/variants/games_api_seed17/tokenizer/chat_template.jinja b/variants/games_api_seed17/tokenizer/chat_template.jinja new file mode 100644 index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4 --- /dev/null +++ b/variants/games_api_seed17/tokenizer/chat_template.jinja @@ -0,0 +1,89 @@ +{%- if tools %} + {{- '<|im_start|>system\n' }} + {%- if messages[0].role == 'system' %} + {{- messages[0].content + '\n\n' }} + {%- endif %} + {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within XML tags:\n" }} + {%- for tool in tools %} + {{- "\n" }} + {{- tool | tojson }} + {%- endfor %} + {{- "\n\n\nFor each function call, return a json object with function name and arguments within XML tags:\n\n{\"name\": , \"arguments\": }\n<|im_end|>\n" }} +{%- else %} + {%- if messages[0].role == 'system' %} + {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }} + {%- endif %} +{%- endif %} +{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %} +{%- for message in messages[::-1] %} + {%- set index = (messages|length - 1) - loop.index0 %} + {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('') and message.content.endswith('')) %} + {%- set ns.multi_step_tool = false %} + {%- set ns.last_query_index = index %} + {%- endif %} +{%- endfor %} +{%- for message in messages %} + {%- if message.content is string %} + {%- set content = message.content %} + {%- else %} + {%- set content = '' %} + {%- endif %} + {%- if (message.role == "user") or (message.role == "system" and not loop.first) %} + {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }} + {%- elif message.role == "assistant" %} + {%- set reasoning_content = '' %} + {%- if message.reasoning_content is string %} + {%- set reasoning_content = message.reasoning_content %} + {%- else %} + {%- if '' in content %} + {%- set reasoning_content = content.split('')[0].rstrip('\n').split('')[-1].lstrip('\n') %} + {%- set content = content.split('')[-1].lstrip('\n') %} + {%- endif %} + {%- endif %} + {%- if loop.index0 > ns.last_query_index %} + {%- if loop.last or (not loop.last and reasoning_content) %} + {{- '<|im_start|>' + message.role + '\n\n' + reasoning_content.strip('\n') + '\n\n\n' + content.lstrip('\n') }} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- if message.tool_calls %} + {%- for tool_call in message.tool_calls %} + {%- if (loop.first and content) or (not loop.first) %} + {{- '\n' }} + {%- endif %} + {%- if tool_call.function %} + {%- set tool_call = tool_call.function %} + {%- endif %} + {{- '\n{"name": "' }} + {{- tool_call.name }} + {{- '", "arguments": ' }} + {%- if tool_call.arguments is string %} + {{- tool_call.arguments }} + {%- else %} + {{- tool_call.arguments | tojson }} + {%- endif %} + {{- '}\n' }} + {%- endfor %} + {%- endif %} + {{- '<|im_end|>\n' }} + {%- elif message.role == "tool" %} + {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %} + {{- '<|im_start|>user' }} + {%- endif %} + {{- '\n\n' }} + {{- content }} + {{- '\n' }} + {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %} + {{- '<|im_end|>\n' }} + {%- endif %} + {%- endif %} +{%- endfor %} +{%- if add_generation_prompt %} + {{- '<|im_start|>assistant\n' }} + {%- if enable_thinking is defined and enable_thinking is false %} + {{- '\n\n\n\n' }} + {%- endif %} +{%- endif %} \ No newline at end of file diff --git a/variants/games_api_seed17/tokenizer/tokenizer.json b/variants/games_api_seed17/tokenizer/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2 --- /dev/null +++ b/variants/games_api_seed17/tokenizer/tokenizer.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506 +size 11422650 diff --git a/variants/games_api_seed17/tokenizer/tokenizer_config.json b/variants/games_api_seed17/tokenizer/tokenizer_config.json new file mode 100644 index 0000000000000000000000000000000000000000..9a87e3822403fb40a9524bb75edb66c6f0585750 --- /dev/null +++ b/variants/games_api_seed17/tokenizer/tokenizer_config.json @@ -0,0 +1,30 @@ +{ + "add_prefix_space": false, + "backend": "tokenizers", + "bos_token": null, + "clean_up_tokenization_spaces": false, + "eos_token": "<|im_end|>", + "errors": "replace", + "extra_special_tokens": [ + "<|im_start|>", + "<|im_end|>", + "<|object_ref_start|>", + "<|object_ref_end|>", + "<|box_start|>", + "<|box_end|>", + "<|quad_start|>", + "<|quad_end|>", + "<|vision_start|>", + "<|vision_end|>", + "<|vision_pad|>", + "<|image_pad|>", + "<|video_pad|>" + ], + "is_local": true, + "local_files_only": true, + "model_max_length": 131072, + "pad_token": "<|endoftext|>", + "split_special_tokens": false, + "tokenizer_class": "Qwen2Tokenizer", + "unk_token": null +} diff --git a/variants/games_api_seed17/train_log.json b/variants/games_api_seed17/train_log.json new file mode 100644 index 0000000000000000000000000000000000000000..3159f3a25d065a4dc48d78c04dbc08f5c0dde01f --- /dev/null +++ b/variants/games_api_seed17/train_log.json @@ -0,0 +1,1445 @@ +[ + { + "step": 1, + "phase": "full", + "loss": 1.0307002067565918, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 1.4512842185795307, + "batch_question_ids_sha256": "b761fda494e336dd1b01f15e1e2340138b46ee1f2f7cdc6e859c47144a82be52" + }, + { + "step": 2, + "phase": "full", + "loss": 0.8930762559175491, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 2.504309923388064, + "batch_question_ids_sha256": "22c7baae42c9d9b79718b5532a042a1fe9ffc6899db2a8cc226dcda5c866b517" + }, + { + "step": 3, + "phase": "full", + "loss": 0.8960195630788803, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 3.670662384480238, + "batch_question_ids_sha256": "aaaf8b0bf19003fef765f486bb32afb0f5a4e09ce0995d8fb106c6bd0a93f599" + }, + { + "step": 4, + "phase": "full", + "loss": 0.9547187685966492, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 4.598304734565318, + "batch_question_ids_sha256": "7d7e52423a416f0b1f10d5c4236dc3abc3bffc7e550e22d889e62c08d65f9432" + }, + { + "step": 5, + "phase": "full", + "loss": 0.7883789241313934, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 5.367147143930197, + "batch_question_ids_sha256": "163044141cac80336b94c75cb51378c081715722cdef9878641d02832e29f355" + }, + { + "step": 6, + "phase": "full", + "loss": 0.768450453877449, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 6.088782921433449, + "batch_question_ids_sha256": "3cafcdf2baa926ce7f1db82fc25c397b0139589682d54294afd28d37d7a766a1" + }, + { + "step": 7, + "phase": "full", + "loss": 0.8187769204378128, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 7.020483755506575, + "batch_question_ids_sha256": "beb635b44f9f478e8b0641986feb8e4b5c1c8137915cd824f262792b39c873cb" + }, + { + "step": 8, + "phase": "full", + "loss": 0.8753918260335922, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 7.814922742545605, + "batch_question_ids_sha256": "08b64178677a605b86f49a29119ea3e9f5b0a48cceb0d0a2e59b53783a3345a4" + }, + { + "step": 9, + "phase": "full", + "loss": 0.7718563377857208, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 8.731685343198478, + "batch_question_ids_sha256": "83f0e09c74b720b837b21f09bd4be78d27ff2c44213024cc99a8fb9679e2c971" + }, + { + "step": 10, + "phase": "full", + "loss": 0.8616485297679901, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 9.708810444921255, + "batch_question_ids_sha256": "a49c69b86c96905fc599b3eefaf381bd2c514ddaf9cf3352bc94c7d46fb42ccf" + }, + { + "step": 11, + "phase": "full", + "loss": 0.6401022374629974, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 10.441212999634445, + "batch_question_ids_sha256": "c28e4b6412916861c3c0c950d7d4ca0df8d141b121a068bf62d4e36eafc4d67b" + }, + { + "step": 12, + "phase": "full", + "loss": 0.9817685335874557, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 11.827318474650383, + "batch_question_ids_sha256": "25f1db6c09802bce1dbc5e47eb34637376fb1d3c66df91ca2fc8ef7505a467bb" + }, + { + "step": 13, + "phase": "full", + "loss": 0.7258476316928864, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 12.859009994193912, + "batch_question_ids_sha256": "7ea725cd46bca87274f92dd90511af90bf9333377f3c2e5e55351e2219b6405b" + }, + { + "step": 14, + "phase": "full", + "loss": 0.5983261615037918, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 13.670102035626769, + "batch_question_ids_sha256": "2f04c049e6af2c31b7b25824ad6076bf892dbb31b577063e87aa63c13aa23eab" + }, + { + "step": 15, + "phase": "full", + "loss": 0.8087402880191803, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 14.698793781921268, + "batch_question_ids_sha256": "d3daaf9eca41da61a02463d19c6614bc5955df971cc0e49f4315488633c39601" + }, + { + "step": 16, + "phase": "full", + "loss": 0.6902796775102615, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 15.646720664575696, + "batch_question_ids_sha256": "398320535fb682db5e65598ee356f2029b0d7557900f345da27405704c4f163c" + }, + { + "step": 17, + "phase": "full", + "loss": 0.7753046154975891, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 16.569391946308315, + "batch_question_ids_sha256": "804d0344c168b6f71f0d1474630c0531e035da4a39aa6910b86b60e841b9096c" + }, + { + "step": 18, + "phase": "full", + "loss": 0.7032596468925476, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 17.423400001600385, + "batch_question_ids_sha256": "30da221cb5f16e811143c618ba3ee797114e9c9570c813b5db3c42f27aea2d0e" + }, + { + "step": 19, + "phase": "full", + "loss": 1.0868135392665863, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 18.75719004869461, + "batch_question_ids_sha256": "c217586e9cc74734b6f61c0ca541af8a12d09334445186ea4003f0a9d5af9710" + }, + { + "step": 20, + "phase": "full", + "loss": 0.824458122253418, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 20.0211059814319, + "batch_question_ids_sha256": "c60c7db3117b307618c85b7830c06735f9f4bf489ebe05dd0610f16060727e71" + }, + { + "step": 21, + "phase": "full", + "loss": 0.8242820203304291, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 21.125790210440755, + "batch_question_ids_sha256": "a38e4e1a56a25f6209e0a1bbd7ca02354efbde4f1f6cb4fdc1c0b82c1bffbc0d" + }, + { + "step": 22, + "phase": "full", + "loss": 0.8063900768756866, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 21.997028869576752, + "batch_question_ids_sha256": "a2430fe5a411b85c90498c3dcf29311510a39616978ea9db10c3a8f205da3bd9" + }, + { + "step": 23, + "phase": "full", + "loss": 0.6933028548955917, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 22.701737948693335, + "batch_question_ids_sha256": "20c4ebdacbf4543140553eae15d2c3e07ae2fc5fc6e69f6d202114d49e932d84" + }, + { + "step": 24, + "phase": "full", + "loss": 0.7520852088928223, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 23.875817344523966, + "batch_question_ids_sha256": "eb4ee0559b5fbb590e9868c9f14b03828b02fff55cb0997b497a4a2332c18bd8" + }, + { + "step": 25, + "phase": "full", + "loss": 0.7319846898317337, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 24.86957229487598, + "batch_question_ids_sha256": "554aa33caaf33436f34c4a7a74c5c3f1dbbca77f9aec9c1d307f8bcbd0c77e12" + }, + { + "step": 26, + "phase": "full", + "loss": 0.7660107240080833, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 25.769649357534945, + "batch_question_ids_sha256": "7bd54967bd461386165d360de7a2cb7d0e18d21a656fec2efc50ef4625834639" + }, + { + "step": 27, + "phase": "full", + "loss": 0.7672473341226578, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 26.821975615806878, + "batch_question_ids_sha256": "5248203737e33806de95bfc35bb75f0e8599286960bc87a2d6243527f8118d97" + }, + { + "step": 28, + "phase": "full", + "loss": 0.8174275457859039, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 27.81877755187452, + "batch_question_ids_sha256": "83ae1ef411017c88d21888fdae8c08cec371facbe3a2a62dae6369e444326f48" + }, + { + "step": 29, + "phase": "full", + "loss": 0.7041522413492203, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 28.677005496807396, + "batch_question_ids_sha256": "01645300c9b01a4389f6c098d83a7029c3b26d1537bf32ecd5ada5a7a04b1235" + }, + { + "step": 30, + "phase": "full", + "loss": 0.730721652507782, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 29.848915762268007, + "batch_question_ids_sha256": "602e7b08a927037783732c56cbcbda2a4171d1408ec4656a8c245ecdad22b8b3" + }, + { + "step": 31, + "phase": "full", + "loss": 0.6919281184673309, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 30.674228763207793, + "batch_question_ids_sha256": "28d9504026c6654779de5bd6cd84e49620e85fe9084107c57611ee69820c96ce" + }, + { + "step": 32, + "phase": "full", + "loss": 0.8873244225978851, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 31.506850712932646, + "batch_question_ids_sha256": "d9c764b738e526ec6bacc087ffa12679bec919ad35596283350f0a1c2af2ab12" + }, + { + "step": 33, + "phase": "full", + "loss": 0.9623820185661316, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 32.56108580157161, + "batch_question_ids_sha256": "808a7b19de3d560b85ad29439565c2b74086c29417c8b7741e1bdc82428595a9" + }, + { + "step": 34, + "phase": "full", + "loss": 0.5770164430141449, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 33.28746292833239, + "batch_question_ids_sha256": "e7e89519564b1f092f9f80466e0d078ebe79e6b4fc931346ed87a6f0b671ac4a" + }, + { + "step": 35, + "phase": "full", + "loss": 0.7978139519691467, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 34.231255498714745, + "batch_question_ids_sha256": "b69bbbced2df42d43452f8081635688d8106ad31b07b2196a60ffddb8c3f3d19" + }, + { + "step": 36, + "phase": "full", + "loss": 0.8947590291500092, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 35.5504483692348, + "batch_question_ids_sha256": "13aa8109d4739866f15429fcc25263ea90cd3508b4fad374dd355c1aac1e6486" + }, + { + "step": 37, + "phase": "full", + "loss": 0.6514073014259338, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 36.36155644245446, + "batch_question_ids_sha256": "9467dfbd23b9b3dcc50ffd25ae79cc03b020328a0f2b5b5cec790f73dadb1b4a" + }, + { + "step": 38, + "phase": "full", + "loss": 0.6856623888015747, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 37.20365169085562, + "batch_question_ids_sha256": "e0af44a3a365a405119e2ea9c6d3b0cedcf8cee14d384479eabf0918cabae4c1" + }, + { + "step": 39, + "phase": "full", + "loss": 0.7830156981945038, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 38.202871021814644, + "batch_question_ids_sha256": "4597617e90848fd5e20745e8757be017d32a0ca61e91b7f6e4ea7403bb9e38df" + }, + { + "step": 40, + "phase": "full", + "loss": 0.8369044810533524, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 39.344195276498795, + "batch_question_ids_sha256": "45e9af11d0584756fb11fd22bda0e62d131946deab11265c2f82df5150b0e07b" + }, + { + "step": 41, + "phase": "full", + "loss": 0.8882690966129303, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 40.75169862154871, + "batch_question_ids_sha256": "960954d2defa480eeb52c30c994a0d02035f80d5f359b3c0b6098a298d948db1" + }, + { + "step": 42, + "phase": "full", + "loss": 0.8227152675390244, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 41.669892501086, + "batch_question_ids_sha256": "d0720e35ea4543fa8503ed58fd68ed7b01a2d346557b7f3aa55423898cddcdfb" + }, + { + "step": 43, + "phase": "full", + "loss": 0.7786084339022636, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 42.81298916041851, + "batch_question_ids_sha256": "72bcb6aaa81962ad09ba4b4a2ba8dec4b2df6ce91f1b908f06c41bf55b275144" + }, + { + "step": 44, + "phase": "full", + "loss": 0.8930060863494873, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 43.785257066600025, + "batch_question_ids_sha256": "ea9c0b60b9308e6bb9d3105de58e1746889c82b8e0cfa2d0928b80569e579ed9" + }, + { + "step": 45, + "phase": "full", + "loss": 0.7108842879533768, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 44.65508048888296, + "batch_question_ids_sha256": "f27fddcdbb7494f8cac848bb212e46e46100387a29453da940b0f7683d73b9b9" + }, + { + "step": 46, + "phase": "full", + "loss": 0.5665369778871536, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 45.4550923993811, + "batch_question_ids_sha256": "259d9f4077aebd13a6b3dc23ce9d2603d75fd5ea2b8d3956cb4f6f8a2ea680b4" + }, + { + "step": 47, + "phase": "full", + "loss": 0.9021382182836533, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 46.80424630083144, + "batch_question_ids_sha256": "16198e6d322e811465ed881b8b131547da71d7a072bb90fec60abc2328224778" + }, + { + "step": 48, + "phase": "full", + "loss": 0.8467951714992523, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 48.164256082847714, + "batch_question_ids_sha256": "2466a322a3663c5f71d64bb29bb03943a01b2288308fd4cef16b818c9b011847" + }, + { + "step": 49, + "phase": "full", + "loss": 0.7896639257669449, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 49.32273670565337, + "batch_question_ids_sha256": "319b2631ac34cee856a10b23d9b8a0adecea9a59cd400dc1ad6b1508ae48a4e7" + }, + { + "step": 50, + "phase": "full", + "loss": 0.804805114865303, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 50.47389311064035, + "batch_question_ids_sha256": "af622d3fac134eadfecefb7588004cfbf1e1ae3178e75ec4af28e3b1a8354fbf", + "dev": { + "questions": 504, + "objective": "teacher", + "eligible_questions": 501, + "target_ce": 0.7503794492244228, + "target_kl": 0.1448544377724438, + "target_tv": 0.1912051424425287, + "by_target_kind": { + "teacher_rounded_proxy": { + "n": 501, + "ce": 0.7503794492244228, + "kl": 0.1448544377724438, + "tv": 0.1912051424425287 + } + }, + "independent_targets": { + "teacher": { + "n": 501, + "ce": 0.7503794492244228, + "kl": 0.1448544377724438, + "tv": 0.1912051424425287 + }, + "gold_distribution": { + "n": 504, + "ce": 0.7033193256847233, + "kl": 0.661593171938471, + "tv": 0.4366130240936953 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + { + "step": 51, + "phase": "full", + "loss": 0.7057921439409256, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 59.07291432656348, + "batch_question_ids_sha256": "af74031f64eb40be701d886fce6c815c3d660c04729655245b0d5afd279798d3" + }, + { + "step": 52, + "phase": "full", + "loss": 0.7136913537979126, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 60.052659389562905, + "batch_question_ids_sha256": "a8be871d08248f5a810e98a9874e3b8f0d10ad06a0c9c7dd6c623f7a13f370dd" + }, + { + "step": 53, + "phase": "full", + "loss": 0.6409988701343536, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 60.749022644013166, + "batch_question_ids_sha256": "7e04858e8b9dde01abd1920aefb2e46c50b937862ed21ddc83d402898f3b6421" + }, + { + "step": 54, + "phase": "full", + "loss": 0.6273103803396225, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 61.70915904361755, + "batch_question_ids_sha256": "a268a5c85dc88431994a4e8705b6a7c21ab3eac355aa43cc3baed3e9e4b76ede" + }, + { + "step": 55, + "phase": "full", + "loss": 0.705372765660286, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 62.80674369446933, + "batch_question_ids_sha256": "6efebca356191db3ee2e1b2f56585a29f3e3cace0d6a9ebf9986bacb6794a17b" + }, + { + "step": 56, + "phase": "full", + "loss": 0.7426139712333679, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 63.66315237246454, + "batch_question_ids_sha256": "6ddff367b83c9b3487b0d94504a62d5c8864d9945c8ed97ea7f496f30248f71a" + }, + { + "step": 57, + "phase": "full", + "loss": 0.6855885088443756, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 64.60096775833517, + "batch_question_ids_sha256": "b15ebad8ba204a67e9b8c286c70e5da2bcb9a5c829aa8f8ff1ddf4e6449bd520" + }, + { + "step": 58, + "phase": "full", + "loss": 0.8127493560314178, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 65.6180268349126, + "batch_question_ids_sha256": "2e98e633d708c7d520fb747816efd8291c4d4b0b88d693ad7863028f651ee45e" + }, + { + "step": 59, + "phase": "full", + "loss": 0.6881939619779587, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 66.56642236374319, + "batch_question_ids_sha256": "1badd56059287e6ff596b73fd6e29223fa275563bcf6d9a8d902652cc77b3368" + }, + { + "step": 60, + "phase": "full", + "loss": 0.5895531177520752, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 67.35980618745089, + "batch_question_ids_sha256": "5893294aa711ea06a09408026645c58330515e07e089b3becb0ebf1610a39eaa" + }, + { + "step": 61, + "phase": "full", + "loss": 0.7612251937389374, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 68.40241683274508, + "batch_question_ids_sha256": "e051addc583f1445f2c25f746f1c4e668f1345585d897fb0b28a37219202ee32" + }, + { + "step": 62, + "phase": "full", + "loss": 0.8217856287956238, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 69.47141242586076, + "batch_question_ids_sha256": "29400a61b3017d7bfe7eb8704308cbf3b18776931c678ac8d67d5a4d67b3f7ab" + }, + { + "step": 63, + "phase": "full", + "loss": 0.6335558742284775, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 70.20863586571068, + "batch_question_ids_sha256": "0eaf00472be5b50a056a0f32085b253b24370fa43667d055801997d7bf981a67" + }, + { + "step": 64, + "phase": "full", + "loss": 0.6322737336158752, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 70.95248525775969, + "batch_question_ids_sha256": "3691ea49626cc8115faf1b860f0d70e3ebeedcfffabc6b5b6e71639c5188cfa3" + }, + { + "step": 65, + "phase": "full", + "loss": 0.7272224724292755, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 71.88478084374219, + "batch_question_ids_sha256": "75c22be8d8f31d240a62223272c0b2cac7c9052aa3983a2a6cc24b23e694f160" + }, + { + "step": 66, + "phase": "full", + "loss": 0.773353099822998, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 72.5951349074021, + "batch_question_ids_sha256": "3e8e9a7d47666507658ae89b525e00063d69b2ac6a30691246ade1db18f9d730" + }, + { + "step": 67, + "phase": "full", + "loss": 0.7414793223142624, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 73.48682408500463, + "batch_question_ids_sha256": "09bea6ffb0b7bf41b019b5f300e496012575d2f897e72da8386a8ed513d6d9c7" + }, + { + "step": 68, + "phase": "full", + "loss": 0.7644875049591064, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 74.4316541729495, + "batch_question_ids_sha256": "a982c9488c3bdff99d03c19fa2ba0c4f368dfca857f15e70ae07563b4432d86e" + }, + { + "step": 69, + "phase": "full", + "loss": 0.7250894457101822, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 75.3860358633101, + "batch_question_ids_sha256": "961b863ba8b308c448e7a914e2ab740aafa64e428549e7550bcec0592ef808d9" + }, + { + "step": 70, + "phase": "full", + "loss": 0.7133085504174232, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 76.16088165994734, + "batch_question_ids_sha256": "13a7c712f50fd82bcdf262428d7e3a7414a64b0f0b368dad42428658ccefaa7d" + }, + { + "step": 71, + "phase": "full", + "loss": 0.8405396044254303, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 77.40507543832064, + "batch_question_ids_sha256": "9ff241aa87123299a9fd966e315ca503035efc8d984cef72637289baa7bb63c6" + }, + { + "step": 72, + "phase": "full", + "loss": 0.5628152340650558, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 78.11989181581885, + "batch_question_ids_sha256": "c6a37d3e359e3e039927c1b212558158094bee20134809c71d7ce946e5421910" + }, + { + "step": 73, + "phase": "full", + "loss": 0.7550799548625946, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 79.05754533596337, + "batch_question_ids_sha256": "dddbcc38226b152ab4495621a9200f83a86408e896c95226e918852293bc0e42" + }, + { + "step": 74, + "phase": "full", + "loss": 0.6896497905254364, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 80.00088995043188, + "batch_question_ids_sha256": "7b4402f910480996ff7071cc2016652be833aa765d756efe3daafc68d6335fbc" + }, + { + "step": 75, + "phase": "full", + "loss": 0.6979458332061768, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 80.84783039987087, + "batch_question_ids_sha256": "e4277e9810226703d48805992a37febe9cac469d37f01942decd508f17b962c1" + }, + { + "step": 76, + "phase": "full", + "loss": 0.7504986971616745, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 81.84340382460505, + "batch_question_ids_sha256": "bfb05e8ca243d2696f92f2e521536ce77ad2d4e4fbca56b46103041da8ff4d7c" + }, + { + "step": 77, + "phase": "full", + "loss": 0.6140897125005722, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 82.69638097472489, + "batch_question_ids_sha256": "90418bb41eae7eaf2b940ec80c4c4498d9e19d86603f970dbda9fcb71b8bd156" + }, + { + "step": 78, + "phase": "full", + "loss": 0.802308663725853, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 83.44441852346063, + "batch_question_ids_sha256": "72eb75c6511b128605d841a40d11dddd186f53d272cb95abe1fb57cc48862c3f" + }, + { + "step": 79, + "phase": "full", + "loss": 0.769298329949379, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 84.48834537155926, + "batch_question_ids_sha256": "592526e6d4fd19dda46ef7f6e03144c8b77a07148cfed30395125b8470c9ebdf" + }, + { + "step": 80, + "phase": "full", + "loss": 0.6242982298135757, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 85.33383495546877, + "batch_question_ids_sha256": "c07fbf6d96830d5742f3c18be3bd9930beea04f813603e034ed30e732734ac53" + }, + { + "step": 81, + "phase": "full", + "loss": 0.6981416940689087, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 86.10189342219383, + "batch_question_ids_sha256": "0765cf7b1efed4fcf2737773738486688a77a8056957144c199a7fb0c6d87a90" + }, + { + "step": 82, + "phase": "full", + "loss": 0.7932993173599243, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 87.1551705589518, + "batch_question_ids_sha256": "2befe1db81dddbece15f311be5913e15a921078733ff181e8605250c9d25dbc2" + }, + { + "step": 83, + "phase": "full", + "loss": 0.6262448951601982, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 87.94073358457536, + "batch_question_ids_sha256": "e46ccb6a19d5cb3a69e966da601cb10bf46ec4c510e9f5d109f38fecdb0a435f" + }, + { + "step": 84, + "phase": "full", + "loss": 0.6279762983322144, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 88.91717649251223, + "batch_question_ids_sha256": "4305e7d859dcc8a3f5fef1395a2873de98f9e5ce2397fed2c0afad16f8ea6441" + }, + { + "step": 85, + "phase": "full", + "loss": 0.6378504186868668, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 89.88755099661648, + "batch_question_ids_sha256": "5889bab54f45f1cd0472ea6942b2cc1fc151b0e1cc242e4a92affb20918872ed" + }, + { + "step": 86, + "phase": "full", + "loss": 0.8101979345083237, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 90.72849850077182, + "batch_question_ids_sha256": "904ab6a0a7d7d5c7cc4067f33df1eac2b46bdaf5990b3266fc8b3176e45cc9af" + }, + { + "step": 87, + "phase": "full", + "loss": 0.714586153626442, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 91.68368878308684, + "batch_question_ids_sha256": "86412e60ff26c8f88a346395a937f0e724e82ea834d0ac51439a24fe868ca81c" + }, + { + "step": 88, + "phase": "full", + "loss": 0.584343820810318, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 92.4582961620763, + "batch_question_ids_sha256": "bad994ee38707fc9e963a3f00262735d951e296ffad384328fa8ce14b930a860" + }, + { + "step": 89, + "phase": "full", + "loss": 0.7040254771709442, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 93.20458764769137, + "batch_question_ids_sha256": "0b10d1f90ce8f5ac91cf269dddad5d29b325b76d38db3ac787f099bdb442d894" + }, + { + "step": 90, + "phase": "full", + "loss": 0.7515884786844254, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 94.15069498401135, + "batch_question_ids_sha256": "05bf1cc6ae8449ab428484c3f1bfc94b535817832906ea75c9a59cabea485487" + }, + { + "step": 91, + "phase": "full", + "loss": 0.5967036038637161, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 94.85291576758027, + "batch_question_ids_sha256": "50f5c52d5b9de0365f62b37aebf2d4cb97e46ae560eb9fa2b4afa7124cce04b2" + }, + { + "step": 92, + "phase": "full", + "loss": 0.7817348092794418, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 95.68817425053567, + "batch_question_ids_sha256": "8bbe31332187628ee597887712b40a6216351f5f848b05deda99c7ff4127bdc1" + }, + { + "step": 93, + "phase": "full", + "loss": 0.6736026406288147, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 96.4736470002681, + "batch_question_ids_sha256": "9188c6bd96ef58723c3c63248ba94a6c0e8cad520ea767216dbac59c6dd60f7e" + }, + { + "step": 94, + "phase": "full", + "loss": 0.6233540624380112, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 97.49574191030115, + "batch_question_ids_sha256": "4ca218181816a314c7bc97a9b95f4f747360af61b412b3a562701a080a0ff03f" + }, + { + "step": 95, + "phase": "full", + "loss": 0.7889072000980377, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 98.47146553173661, + "batch_question_ids_sha256": "ecb8c2146c63baf76ecf0ef15f76312f519707d2b53cc8654c0d217522b87d0a" + }, + { + "step": 96, + "phase": "full", + "loss": 0.6907715797424316, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 99.3967102933675, + "batch_question_ids_sha256": "aaf68ecc60a1b646adb51e8fc725ec602a3b83a658cdb5b5663a3a45f7fbade7" + }, + { + "step": 97, + "phase": "full", + "loss": 0.6606822609901428, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 100.26167400646955, + "batch_question_ids_sha256": "cc060c4151a007f17da81a2c7c836f70092b23bcda10d30c2e4f6257f08bb941" + }, + { + "step": 98, + "phase": "full", + "loss": 0.6383478567004204, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 100.9843971291557, + "batch_question_ids_sha256": "225013a0c5f55a80abf9ed69be76bfd4f8066595f1bd457fa792f87033941941" + }, + { + "step": 99, + "phase": "full", + "loss": 0.6590020060539246, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 101.76605395972729, + "batch_question_ids_sha256": "8c67818ea12555d6f525e6f26d4d7153135b386093ac509bea46cc8d7f4c3f96" + }, + { + "step": 100, + "phase": "full", + "loss": 0.7614487111568451, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 102.63410064857453, + "batch_question_ids_sha256": "6e3a28926f93754ccff5e7c7267b7118dacd3fbc5a3cbcc621e7045a217596f0", + "dev": { + "questions": 504, + "objective": "teacher", + "eligible_questions": 501, + "target_ce": 0.733226254728955, + "target_kl": 0.12770124327697543, + "target_tv": 0.1730296882782182, + "by_target_kind": { + "teacher_rounded_proxy": { + "n": 501, + "ce": 0.733226254728955, + "kl": 0.12770124327697543, + "tv": 0.1730296882782182 + } + }, + "independent_targets": { + "teacher": { + "n": 501, + "ce": 0.733226254728955, + "kl": 0.12770124327697543, + "tv": 0.1730296882782182 + }, + "gold_distribution": { + "n": 504, + "ce": 0.7100561694137759, + "kl": 0.6683300156675238, + "tv": 0.4300706943061034 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + { + "step": 101, + "phase": "full", + "loss": 0.8137390911579132, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 111.58826769050211, + "batch_question_ids_sha256": "23dcfc0893870830b3b6e45ba85ac0367e6563e1b32f988eedfe094ad3304cf8" + }, + { + "step": 102, + "phase": "full", + "loss": 0.6275766715407372, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 112.4538603425026, + "batch_question_ids_sha256": "3c3159608817ee81cf0b1d733aa7564d2ba40280be23e667a3bcebf37699e928" + }, + { + "step": 103, + "phase": "full", + "loss": 0.6978338807821274, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 113.16733362432569, + "batch_question_ids_sha256": "3ab2cf8ddf999144adf7686659cd279e4cd1b9cfaa9e5b4b8c9f541bf1730b66" + }, + { + "step": 104, + "phase": "full", + "loss": 0.6379766464233398, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 113.87387787085027, + "batch_question_ids_sha256": "2d6fd4531f6969777d38fa004062dc11035b2a1f4744577fc69ac06d1a88e639" + }, + { + "step": 105, + "phase": "full", + "loss": 0.6705674082040787, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 114.60576388705522, + "batch_question_ids_sha256": "0a840e80c85421ede989b31d353c6936aee18c6ca94ca0d6010391a2513c994f" + }, + { + "step": 106, + "phase": "full", + "loss": 0.7481615990400314, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 115.37257293052971, + "batch_question_ids_sha256": "8b65c4c2a9f39fdf7526a536da01c7c54200d84f35a2723a613d2eea4999461a" + }, + { + "step": 107, + "phase": "full", + "loss": 0.6470892429351807, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 116.27892713528126, + "batch_question_ids_sha256": "79ecea658371bdc56d3863d6a8870392d542d8224d4fa76f36f755a669e3e8ec" + }, + { + "step": 108, + "phase": "full", + "loss": 0.7184521853923798, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 117.07777258660644, + "batch_question_ids_sha256": "c66681e1ab70c48e2a76bbf661bc8439d7f4f157cc2b6172e1d94fb34e3c80ab" + }, + { + "step": 109, + "phase": "full", + "loss": 0.8457076698541641, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 118.51581192482263, + "batch_question_ids_sha256": "2e4d0f48e411d074b89a3cbc774e4db5ebf5380a17a75ed451bf6047565f6869" + }, + { + "step": 110, + "phase": "full", + "loss": 0.8735138326883316, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 119.90560970734805, + "batch_question_ids_sha256": "1711f23c93a289ae28373c19c23d30fdc281e832512651cddc2ef5ba1df6de22" + }, + { + "step": 111, + "phase": "full", + "loss": 0.6799501478672028, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 120.71871845051646, + "batch_question_ids_sha256": "5dbdc92a859ac400fe77da667326319af37cdcc43841585594ab3662774be01a" + }, + { + "step": 112, + "phase": "full", + "loss": 0.6167872101068497, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 121.46985782403499, + "batch_question_ids_sha256": "fa91abcbc68a50f5fcec4ba70494ce5600b81862fed30ae0edbecd46bd531645" + }, + { + "step": 113, + "phase": "full", + "loss": 0.5517106503248215, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 122.24117500148714, + "batch_question_ids_sha256": "43083db20364fc2606168823acb85a82e1c059f0eced4ef8751b1f2050429a9d" + }, + { + "step": 114, + "phase": "full", + "loss": 0.6836997270584106, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 123.14003321249038, + "batch_question_ids_sha256": "f297975891671365963ac6c0e517b015f266b64588ea3cfa9feb8f84880c1cae" + }, + { + "step": 115, + "phase": "full", + "loss": 0.618137001991272, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 123.99673863034695, + "batch_question_ids_sha256": "a4504dcb6eacf7f576f330343103903d41e6153602f57836b45cf76fad6e4a9c" + }, + { + "step": 116, + "phase": "full", + "loss": 0.6727172434329987, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 124.790079892613, + "batch_question_ids_sha256": "faf78d82a87994597517ad9c50462cf190bd2ffa102605380cc2f97eb756050a" + }, + { + "step": 117, + "phase": "full", + "loss": 0.6768207997083664, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 125.50653744768351, + "batch_question_ids_sha256": "b345e77e1330589a20d44403b17eb323df21917e534b2f3ef115b63881db210c" + }, + { + "step": 118, + "phase": "full", + "loss": 0.6197745949029922, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 126.51552114635706, + "batch_question_ids_sha256": "706af5a245429c43c3e729dcb1d1e14f9e85aaf898ff49ebb4d783b95189f5af" + }, + { + "step": 119, + "phase": "full", + "loss": 0.6069883853197098, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 127.37185926735401, + "batch_question_ids_sha256": "36234acc4d3ff448155ba98593b83cff76153f95e42a77a841a354badd9ce245" + }, + { + "step": 120, + "phase": "full", + "loss": 0.7446332275867462, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 128.41723258327693, + "batch_question_ids_sha256": "60a4810706d76b7e2dc0327e2e525f760a39253887ad14435adcd0ef2905e471" + }, + { + "step": 121, + "phase": "full", + "loss": 0.6242263317108154, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 129.15088483132422, + "batch_question_ids_sha256": "9b216ff0c733e0fa41d319d48ca51b7bd08e1137a4ec5c68a006672d881a9b73" + }, + { + "step": 122, + "phase": "full", + "loss": 0.6025591790676117, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 129.90009909123182, + "batch_question_ids_sha256": "0ff0ffce5ca0d51206a0f97c261ef785c1da283bf59af809d5451d4c1fa0c1ea" + }, + { + "step": 123, + "phase": "full", + "loss": 0.703130230307579, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 130.80700931232423, + "batch_question_ids_sha256": "77457bb7a51b55ce555789e9ae0ef8874c40e861196269411996290c40f56270" + }, + { + "step": 124, + "phase": "full", + "loss": 0.630307599902153, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 131.6853904435411, + "batch_question_ids_sha256": "ceb6c3368fc2f163c1e697f9f408d6f4b2cd8393e39b3e1fb2652a4c30b434c9" + }, + { + "step": 125, + "phase": "full", + "loss": 0.6928221732378006, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 132.6631599245593, + "batch_question_ids_sha256": "ebbaf818883a14cfa4f9c5bd9fc980ac24b3e11dded0005a66fc870efa6b3798" + }, + { + "step": 126, + "phase": "full", + "loss": 0.746506929397583, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 133.6392150335014, + "batch_question_ids_sha256": "125a1bf29bdd40000d8c79458238d652fa9907962f90606d155031b0ab1a7415" + }, + { + "step": 127, + "phase": "full", + "loss": 0.6362086683511734, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 134.59868404828012, + "batch_question_ids_sha256": "fe0bccd9f0fcad0397d8b40dd04647365248ab2ca2229f0e34d1f9a339837337" + }, + { + "step": 128, + "phase": "full", + "loss": 0.6567190736532211, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 135.51323157921433, + "batch_question_ids_sha256": "11d70ade9ef0bb5fe1c1b35e296d4e19473d40d5711eb677634dd5359598da99" + }, + { + "step": 129, + "phase": "full", + "loss": 0.5145155638456345, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 136.21974855288863, + "batch_question_ids_sha256": "706a450c79a3b24f10fd6f99f21ab696e528c810c1fd7e791ad08b70ae4328d7" + }, + { + "step": 130, + "phase": "full", + "loss": 0.7381799221038818, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 137.1286107506603, + "batch_question_ids_sha256": "3171ccce13ca11f4be88a2243d78aed635c8fe6db83ea6bf76b8e88d97092691" + }, + { + "step": 131, + "phase": "full", + "loss": 0.749908372759819, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 138.2193145994097, + "batch_question_ids_sha256": "7b6f5ff6f8b1675e0798926480d74ea46b20edb6b9458f3b2e04dc3d221fa6be" + }, + { + "step": 132, + "phase": "full", + "loss": 0.6708684861660004, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 138.94129990134388, + "batch_question_ids_sha256": "d308975bd79faf473f10f47207885dea846ac69c124fa0d348be99afa3bac883" + }, + { + "step": 133, + "phase": "full", + "loss": 0.68851737678051, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 139.99681051354855, + "batch_question_ids_sha256": "d6e2c91282e0741cf025c00c88cd6a9b5f2a437b1e578307e7fd1ec353c2061c" + }, + { + "step": 134, + "phase": "full", + "loss": 0.6098454296588898, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 140.75884068664163, + "batch_question_ids_sha256": "117ac91f3b2f0de29991344b996969b01004723483e3455cd864a8066329b635" + }, + { + "step": 135, + "phase": "full", + "loss": 0.6223083510994911, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 141.6828871369362, + "batch_question_ids_sha256": "a93892fdc00ae8158c9c34a7b2ced9863cd44d34d4a6d69a2df587ec7deb549a" + }, + { + "step": 136, + "phase": "full", + "loss": 0.6275051683187485, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 142.44840356428176, + "batch_question_ids_sha256": "002ca28ddc7d88b6c050ec76dd9bd1fa94270172cf65ed9e94587a61e15fd4d4" + }, + { + "step": 137, + "phase": "full", + "loss": 0.6827020645141602, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 143.44316311739385, + "batch_question_ids_sha256": "43cead99a9912e08a2275677b0daee3d1eb48e76fa6e6ec3bfedd5209490248f" + }, + { + "step": 138, + "phase": "full", + "loss": 0.7526485770940781, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 144.14391784649342, + "batch_question_ids_sha256": "7edb9ac12e29f5229cffbfe76612f710f3e0a49dca15c4122640876f0633ab01" + }, + { + "step": 139, + "phase": "full", + "loss": 0.5887515395879745, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 145.07921190466732, + "batch_question_ids_sha256": "474395846b81143c53dcb4fd1d0d704d4ac24d8b29d975171a6477f74b03b87f" + }, + { + "step": 140, + "phase": "full", + "loss": 0.7421691417694092, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 146.04882868751884, + "batch_question_ids_sha256": "164e5fe7cf5f218fb15f912af90dbc455a04968869ee9c8c46b1875fa669d44f" + }, + { + "step": 141, + "phase": "full", + "loss": 0.6369046419858932, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 146.8598380368203, + "batch_question_ids_sha256": "dbd15c584e3a6c840d13143252d1325d81979543e8c701dc3147eed8fe5d9ba2" + }, + { + "step": 142, + "phase": "full", + "loss": 0.592265784740448, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 147.54976333584636, + "batch_question_ids_sha256": "15e70742d9af90d4793f9019fd9ff6889ada836af6ac23ff5ca6855c297d5a52" + }, + { + "step": 143, + "phase": "full", + "loss": 0.7309145480394363, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 148.26254807133228, + "batch_question_ids_sha256": "bb0d8460122b4250011b3f6f3a42f5a63525b6863d1ab4cdb5a52466afc72ca5" + }, + { + "step": 144, + "phase": "full", + "loss": 0.6246269047260284, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 149.0716409943998, + "batch_question_ids_sha256": "249a7bd9835453cb78243868d2ec390cb30facda1b0286d0c2e2f8916d15464c" + }, + { + "step": 145, + "phase": "full", + "loss": 0.7877713590860367, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 150.13167964853346, + "batch_question_ids_sha256": "74b7e541c82d66491a34c07c98a9e38ea9d336989d38cf1a84275c2140b4b24b" + }, + { + "step": 146, + "phase": "full", + "loss": 0.718550369143486, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 151.0788530986756, + "batch_question_ids_sha256": "d06298f404a2ab748db4e5ff76216dbe4403456c43f7f476400e2f5387aa2dc5" + }, + { + "step": 147, + "phase": "full", + "loss": 0.7214168310165405, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 151.98923973552883, + "batch_question_ids_sha256": "0e4236760362e9603f0257d7af1890fda6e102514ace2127641727e905e4090c" + }, + { + "step": 148, + "phase": "full", + "loss": 0.6426246166229248, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 152.9279158245772, + "batch_question_ids_sha256": "5262b95d34b8fb8dd8e046d98fbcd751a9b14d3320fb8bac3190f7f522a56062" + }, + { + "step": 149, + "phase": "full", + "loss": 0.7218716740608215, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 153.92151635233313, + "batch_question_ids_sha256": "256afd40232cb1c3e00d73481a5bb782d1f045fb1d355b99676baf3bf3783f47" + }, + { + "step": 150, + "phase": "full", + "loss": 0.613991841673851, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 154.74634620919824, + "batch_question_ids_sha256": "652bbac90b35683f612214243594cc9cc98d2095313111ccd4f33275c9a8c5a8", + "dev": { + "questions": 504, + "objective": "teacher", + "eligible_questions": 501, + "target_ce": 0.7130707709730839, + "target_kl": 0.10754575952110457, + "target_tv": 0.1622813057690467, + "by_target_kind": { + "teacher_rounded_proxy": { + "n": 501, + "ce": 0.7130707709730839, + "kl": 0.10754575952110457, + "tv": 0.1622813057690467 + } + }, + "independent_targets": { + "teacher": { + "n": 501, + "ce": 0.7130707709730839, + "kl": 0.10754575952110457, + "tv": 0.1622813057690467 + }, + "gold_distribution": { + "n": 504, + "ce": 0.701963093111963, + "kl": 0.6602369393657108, + "tv": 0.4357824048845307 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + } +] diff --git a/variants/games_gold_seed17/backbone_config/config.json b/variants/games_gold_seed17/backbone_config/config.json new file mode 100644 index 0000000000000000000000000000000000000000..b815be77563f3eef88baf4b4c8ae59706e618ad2 --- /dev/null +++ b/variants/games_gold_seed17/backbone_config/config.json @@ -0,0 +1,63 @@ +{ + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "bos_token_id": 151643, + "dtype": "float32", + "eos_token_id": 151645, + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 40960, + "max_window_layers": 28, + "model_type": "qwen3", + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "pad_token_id": null, + "rms_norm_eps": 1e-06, + "rope_parameters": { + "rope_theta": 1000000, + "rope_type": "default" + }, + "sliding_window": null, + "tie_word_embeddings": true, + "transformers_version": "5.17.0", + "use_cache": false, + "use_sliding_window": false, + "vocab_size": 151936 +} diff --git a/variants/games_gold_seed17/best.safetensors b/variants/games_gold_seed17/best.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..7bfba464356416bcce2aa2cc1c05e33940c1f9e5 --- /dev/null +++ b/variants/games_gold_seed17/best.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a395c56fad2a9bd4fb664ef78e1e06c19102886d97dd56dfe2676e5896c7e3a +size 2385039280 diff --git a/variants/games_gold_seed17/config.json b/variants/games_gold_seed17/config.json new file mode 100644 index 0000000000000000000000000000000000000000..1c480b24b6b005dd24f96106533e82eba53972d7 --- /dev/null +++ b/variants/games_gold_seed17/config.json @@ -0,0 +1,69 @@ +{ + "input": "data/scaled_games_v4b/policy", + "output_dir": "runs/games_gold_seed17", + "model": "Qwen/Qwen3-0.6B", + "revision": "main", + "init_checkpoint": "/home/rwang/openjev_codex_20260917/runs/v3_teacher_coords_multi_seed17", + "objective": "gold_distribution", + "loss": "ce", + "reward_samples": 32, + "gradient_checkpointing": true, + "set_head": "attention", + "steps": 150, + "head_steps": 0, + "batch_questions": 16, + "microbatch_questions": 4, + "max_microbatch_tokens": 16384, + "eval_every": 50, + "max_length": 8192, + "seed": 17, + "backbone_lr": 2e-05, + "head_lr": 0.0002, + "head_warmup_lr": 0.001, + "precision": "bf16", + "disable_native_triton": true, + "validate_only": false, + "self_check": false, + "schema_version": "openjev-decision-pipeline-v1", + "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "initialization": "local DecisionModel warm start, fresh optimizer", + "data_sha256": { + "data/scaled_games_v4b/policy/train.jsonl": "af6065315d383bce2f137b3095cc9912aafe1edf0a3ee7c119c9b7af4a0ac399", + "data/scaled_games_v4b/policy/dev.jsonl": "5e34f86e60dffce37d9454c4ffb978c987ba8467de35c5f79dab99bdb24f3d6b", + "data/scaled_games_v4b/policy/calibration.jsonl": "c2b279ecd835068c20bc328c34bca970feee1e8afab67b1f7dea0c201f0d824f", + "data/scaled_games_v4b/policy/test.jsonl": "54e573515e5d1e2c55308763a857d9bb46f2c02b2cbd23e50a4544587a69501c", + "data/scaled_games_v4b/policy/ood.jsonl": "ac4ea92590a95f40cd2134ad64f450c369601745b36d91ebf5692d9f39ea3135" + }, + "deps": { + "torch": "2.14.0", + "transformers": "5.17.0", + "safetensors": "0.8.0" + }, + "gpu": "NVIDIA A100-SXM4-80GB", + "parameter_storage": "float32", + "forward_autocast": "bfloat16", + "parameter_count": 596250498, + "train_questions": 1531, + "all_train_questions": 1531, + "schema_counts": { + "records": 589, + "questions_by_split": { + "train": 1531, + "dev": 504, + "calibration": 500, + "test": 487, + "ood": 173 + }, + "eligible_by_split_objective": { + "calibration/gold_distribution": 500, + "dev/gold_distribution": 504, + "ood/gold_distribution": 173, + "test/gold_distribution": 487, + "train/gold_distribution": 1531 + } + }, + "loss_description": "per-question proper loss; complete candidate normalization", + "rlcd_candidate": null, + "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection", + "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing" +} diff --git a/variants/games_gold_seed17/summary.json b/variants/games_gold_seed17/summary.json new file mode 100644 index 0000000000000000000000000000000000000000..3536e2e7fa97b9522ce3d4c6490ea1189bec6ef1 --- /dev/null +++ b/variants/games_gold_seed17/summary.json @@ -0,0 +1,160 @@ +{ + "best_step": 100, + "best_dev_target_ce": 0.6079203053015539, + "selected_on": "dev target CE", + "objective": "gold_distribution", + "metrics_by_split": { + "dev": { + "questions": 504, + "objective": "gold_distribution", + "eligible_questions": 504, + "target_ce": 0.6079203053015539, + "target_kl": 0.566194151555302, + "target_tv": 0.33998366396415136, + "by_target_kind": { + "deterministic_truth": { + "n": 423, + "ce": 0.5475351556056708, + "kl": 0.5475351556056708, + "tv": 0.3248628333506329 + }, + "optimal_action_policy": { + "n": 81, + "ce": 0.9232649759356132, + "kl": 0.6636355748478214, + "tv": 0.41894800161253004 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 504, + "ce": 0.6079203053015539, + "kl": 0.566194151555302, + "tv": 0.33998366396415136 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "calibration": { + "questions": 500, + "objective": "gold_distribution", + "eligible_questions": 500, + "target_ce": 0.6601248196618281, + "target_kl": 0.6225635378425565, + "target_tv": 0.35279715502465236, + "by_target_kind": { + "deterministic_truth": { + "n": 420, + "ce": 0.6192864014688629, + "kl": 0.6192864014688629, + "tv": 0.34447223537915905 + }, + "optimal_action_policy": { + "n": 80, + "ce": 0.8745265151748954, + "kl": 0.6397685038044467, + "tv": 0.39650298316349475 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 500, + "ce": 0.6601248196618281, + "kl": 0.6225635378425565, + "tv": 0.35279715502465236 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "test": { + "questions": 487, + "objective": "gold_distribution", + "eligible_questions": 487, + "target_ce": 0.6302541688123193, + "target_kl": 0.5895691869564941, + "target_tv": 0.33474409936735405, + "by_target_kind": { + "optimal_action_policy": { + "n": 82, + "ce": 0.904787119991967, + "kl": 0.6631580204335937, + "tv": 0.40845001808807946 + }, + "deterministic_truth": { + "n": 405, + "ce": 0.5746697194376736, + "kl": 0.5746697194376736, + "tv": 0.3198209257004425 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 487, + "ce": 0.6302541688123193, + "kl": 0.5895691869564941, + "tv": 0.33474409936735405 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "ood": { + "questions": 173, + "objective": "gold_distribution", + "eligible_questions": 173, + "target_ce": 0.5910269532386467, + "target_kl": 0.5486169132750188, + "target_tv": 0.3325766694434778, + "by_target_kind": { + "deterministic_truth": { + "n": 144, + "ce": 0.5162414440562704, + "kl": 0.5162414440562704, + "tv": 0.31279073019126086 + }, + "optimal_action_policy": { + "n": 29, + "ce": 0.9623756884890643, + "kl": 0.7093778638784569, + "tv": 0.43082409193724464 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 173, + "ce": 0.5910269532386467, + "kl": 0.5486169132750188, + "tv": 0.3325766694434778 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + "temperature": 1.0, + "temperature_fitted": false, + "training_seconds": 186.51977808214724, + "max_gpu_allocated_gb": 12.369258496 +} diff --git a/variants/games_gold_seed17/target_audit.json b/variants/games_gold_seed17/target_audit.json new file mode 100644 index 0000000000000000000000000000000000000000..9d515389414571e89e08198299a113f04405c13d --- /dev/null +++ b/variants/games_gold_seed17/target_audit.json @@ -0,0 +1,54317 @@ +[ + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 151, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 132, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:45a206dbc4e330f637ece6ae:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:45a206dbc4e330f637ece6ae:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:45a206dbc4e330f637ece6ae:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:45a206dbc4e330f637ece6ae:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8d468090944737c21080f821:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8d468090944737c21080f821:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8d468090944737c21080f821:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8d468090944737c21080f821:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5d98b0d63d0ed94d96c88e17:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5d98b0d63d0ed94d96c88e17:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5d98b0d63d0ed94d96c88e17:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5d98b0d63d0ed94d96c88e17:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:da9d9f1b68c04ca029a5e799:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:da9d9f1b68c04ca029a5e799:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:da9d9f1b68c04ca029a5e799:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:da9d9f1b68c04ca029a5e799:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58d3c0dd72c2d2cd11756296:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58d3c0dd72c2d2cd11756296:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58d3c0dd72c2d2cd11756296:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58d3c0dd72c2d2cd11756296:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:07fcac0e43c5e1b61725bfd5:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:07fcac0e43c5e1b61725bfd5:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:07fcac0e43c5e1b61725bfd5:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:07fcac0e43c5e1b61725bfd5:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:16788c902da368b55922674d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:16788c902da368b55922674d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:16788c902da368b55922674d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:16788c902da368b55922674d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1ba43d0958ba5992603084b9:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1ba43d0958ba5992603084b9:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1ba43d0958ba5992603084b9:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1ba43d0958ba5992603084b9:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:bb33abaff323aa42615de283:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:bb33abaff323aa42615de283:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:bb33abaff323aa42615de283:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:bb33abaff323aa42615de283:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:410f76de8b291ce18af3283b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 188, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:410f76de8b291ce18af3283b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:410f76de8b291ce18af3283b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:410f76de8b291ce18af3283b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3a37c45b8893e443bb48d780:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 188, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3a37c45b8893e443bb48d780:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3a37c45b8893e443bb48d780:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3a37c45b8893e443bb48d780:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 155, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 155, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 155, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 155, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4c9ad6e745c09cd665f24370:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4c9ad6e745c09cd665f24370:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4c9ad6e745c09cd665f24370:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4c9ad6e745c09cd665f24370:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3e21643fee7f60ac15597733:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3e21643fee7f60ac15597733:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3e21643fee7f60ac15597733:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3e21643fee7f60ac15597733:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dc47b4d49347923b0a9e3af6:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dc47b4d49347923b0a9e3af6:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dc47b4d49347923b0a9e3af6:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dc47b4d49347923b0a9e3af6:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:adbaaf8a521d504bf37f192d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:adbaaf8a521d504bf37f192d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:adbaaf8a521d504bf37f192d:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:adbaaf8a521d504bf37f192d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 129, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 149, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 130, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 148, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 129, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 149, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 130, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:83f43ddfcbd1718e7a64288e:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:83f43ddfcbd1718e7a64288e:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:83f43ddfcbd1718e7a64288e:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:83f43ddfcbd1718e7a64288e:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5b6fc4a98ac6c78a118ac241:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5b6fc4a98ac6c78a118ac241:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5b6fc4a98ac6c78a118ac241:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5b6fc4a98ac6c78a118ac241:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:caf8e8d695dc780f1b381231:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:caf8e8d695dc780f1b381231:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:caf8e8d695dc780f1b381231:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:caf8e8d695dc780f1b381231:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:881dea5763a938372f85c67c:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:881dea5763a938372f85c67c:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:881dea5763a938372f85c67c:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:881dea5763a938372f85c67c:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 130, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 164, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 148, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 129, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e04aa1b9b164091e5b605d77:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e04aa1b9b164091e5b605d77:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e04aa1b9b164091e5b605d77:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e04aa1b9b164091e5b605d77:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:07fc2dce6ac4c6390adfb38b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:07fc2dce6ac4c6390adfb38b:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:07fc2dce6ac4c6390adfb38b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:07fc2dce6ac4c6390adfb38b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:cf4d4a371cc2b84d602a4bb9:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:cf4d4a371cc2b84d602a4bb9:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:cf4d4a371cc2b84d602a4bb9:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:cf4d4a371cc2b84d602a4bb9:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ebfaca907fc41ff93ba56e2d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ebfaca907fc41ff93ba56e2d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ebfaca907fc41ff93ba56e2d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ebfaca907fc41ff93ba56e2d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 132, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5d4c1f3be8e78fae1f7c577b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5d4c1f3be8e78fae1f7c577b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5d4c1f3be8e78fae1f7c577b:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5d4c1f3be8e78fae1f7c577b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:49cba9d9430dfafd8026f045:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:49cba9d9430dfafd8026f045:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:49cba9d9430dfafd8026f045:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:49cba9d9430dfafd8026f045:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:89f4318abbbf35dae0a9d5fc:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:89f4318abbbf35dae0a9d5fc:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:89f4318abbbf35dae0a9d5fc:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:89f4318abbbf35dae0a9d5fc:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 144, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 158, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 145, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 158, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 145, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 158, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 145, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d1757a09cb988028246a2278:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d1757a09cb988028246a2278:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d1757a09cb988028246a2278:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d1757a09cb988028246a2278:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1f94c03e232d5cecf420b4b0:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1f94c03e232d5cecf420b4b0:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1f94c03e232d5cecf420b4b0:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1f94c03e232d5cecf420b4b0:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9bc6c48d7b6609943aa22be4:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9bc6c48d7b6609943aa22be4:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9bc6c48d7b6609943aa22be4:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9bc6c48d7b6609943aa22be4:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 156, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 143, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 157, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 144, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 155, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0994b5c61a691e7ea285be3f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0994b5c61a691e7ea285be3f:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0994b5c61a691e7ea285be3f:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0994b5c61a691e7ea285be3f:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:cfd599b961f4135d3c87387f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:cfd599b961f4135d3c87387f:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:cfd599b961f4135d3c87387f:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:cfd599b961f4135d3c87387f:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:39b76b489b61357e661386a2:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:39b76b489b61357e661386a2:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:39b76b489b61357e661386a2:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:39b76b489b61357e661386a2:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3643892fdb0609a4df6b640d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3643892fdb0609a4df6b640d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3643892fdb0609a4df6b640d:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3643892fdb0609a4df6b640d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f0189682c2eb6935a5d15fe8:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f0189682c2eb6935a5d15fe8:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f0189682c2eb6935a5d15fe8:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f0189682c2eb6935a5d15fe8:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:981ca032260a3dffcec1bf99:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:981ca032260a3dffcec1bf99:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:981ca032260a3dffcec1bf99:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:981ca032260a3dffcec1bf99:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:373cfe9faa2058b0490d322a:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:373cfe9faa2058b0490d322a:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:373cfe9faa2058b0490d322a:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:373cfe9faa2058b0490d322a:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:16756b2a682710731d258d4b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:16756b2a682710731d258d4b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:16756b2a682710731d258d4b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:16756b2a682710731d258d4b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:696202fd3a2ce99a2445985e:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:696202fd3a2ce99a2445985e:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:696202fd3a2ce99a2445985e:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:696202fd3a2ce99a2445985e:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a9cd87f6087d543fb114185:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a9cd87f6087d543fb114185:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a9cd87f6087d543fb114185:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a9cd87f6087d543fb114185:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 156, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 143, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 157, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 144, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 155, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 157, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 144, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a1a646f11d996660cf23de69:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a1a646f11d996660cf23de69:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a1a646f11d996660cf23de69:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a1a646f11d996660cf23de69:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:54d4c05b4976562ce121bf21:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:54d4c05b4976562ce121bf21:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:54d4c05b4976562ce121bf21:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:54d4c05b4976562ce121bf21:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9d823e17bb4ea3f97c12577e:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9d823e17bb4ea3f97c12577e:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9d823e17bb4ea3f97c12577e:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9d823e17bb4ea3f97c12577e:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:56e8bd28a696b0a31a95926f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:56e8bd28a696b0a31a95926f:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:56e8bd28a696b0a31a95926f:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:56e8bd28a696b0a31a95926f:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 225, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 253, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 253, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 253, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 253, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 224, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 237, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 224, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 252, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1a3e61f06ce2e98ebb505ccb:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1a3e61f06ce2e98ebb505ccb:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1a3e61f06ce2e98ebb505ccb:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1a3e61f06ce2e98ebb505ccb:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e4df2f7dacb5671625749f3e:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e4df2f7dacb5671625749f3e:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e4df2f7dacb5671625749f3e:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e4df2f7dacb5671625749f3e:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:edf573117fe7de0a60f09ec4:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 195, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:edf573117fe7de0a60f09ec4:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:edf573117fe7de0a60f09ec4:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:edf573117fe7de0a60f09ec4:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f3ded4511a05512af9c015ed:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 186, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f3ded4511a05512af9c015ed:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f3ded4511a05512af9c015ed:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f3ded4511a05512af9c015ed:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 197, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 219, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 197, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:63eb6eaccebe64181ab30b1a:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:63eb6eaccebe64181ab30b1a:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:63eb6eaccebe64181ab30b1a:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:63eb6eaccebe64181ab30b1a:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:19d04caecfe12a6a9c24e521:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:19d04caecfe12a6a9c24e521:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:19d04caecfe12a6a9c24e521:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:19d04caecfe12a6a9c24e521:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5eea186d00c75f1dfa595738:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5eea186d00c75f1dfa595738:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5eea186d00c75f1dfa595738:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5eea186d00c75f1dfa595738:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:067ceb1f04a6877293ed2723:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:067ceb1f04a6877293ed2723:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:067ceb1f04a6877293ed2723:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:067ceb1f04a6877293ed2723:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 227, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 224, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d176ad19885d837f1a7982fa:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d176ad19885d837f1a7982fa:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d176ad19885d837f1a7982fa:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d176ad19885d837f1a7982fa:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0efeba2cbe230918df7006fa:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0efeba2cbe230918df7006fa:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0efeba2cbe230918df7006fa:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0efeba2cbe230918df7006fa:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:94c72585efdea8dab81e3ed6:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:94c72585efdea8dab81e3ed6:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:94c72585efdea8dab81e3ed6:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:94c72585efdea8dab81e3ed6:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e1e15ce94bc3c638ab2c3eb0:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e1e15ce94bc3c638ab2c3eb0:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e1e15ce94bc3c638ab2c3eb0:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e1e15ce94bc3c638ab2c3eb0:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 235, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 235, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 236, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 235, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:72dfb5d2d802b721e1e7ce85:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:72dfb5d2d802b721e1e7ce85:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:72dfb5d2d802b721e1e7ce85:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:72dfb5d2d802b721e1e7ce85:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5e8e60860af25ef0e831dd82:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5e8e60860af25ef0e831dd82:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5e8e60860af25ef0e831dd82:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5e8e60860af25ef0e831dd82:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:96fc807e107c37d4a70d690a:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:96fc807e107c37d4a70d690a:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:96fc807e107c37d4a70d690a:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:96fc807e107c37d4a70d690a:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 199, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 219, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 219, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:89f007044a81d170cbb9af1c:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:89f007044a81d170cbb9af1c:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:89f007044a81d170cbb9af1c:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:89f007044a81d170cbb9af1c:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dc616b22757bbd0b1df4e3c1:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dc616b22757bbd0b1df4e3c1:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dc616b22757bbd0b1df4e3c1:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dc616b22757bbd0b1df4e3c1:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e81d37b1cc0f0c63ff630476:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e81d37b1cc0f0c63ff630476:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e81d37b1cc0f0c63ff630476:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e81d37b1cc0f0c63ff630476:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:2569997e97622b8fb7b847f3:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 186, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:2569997e97622b8fb7b847f3:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:2569997e97622b8fb7b847f3:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:2569997e97622b8fb7b847f3:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 227, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 227, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 226, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7c028b79a76c86db17e4e355:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7c028b79a76c86db17e4e355:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7c028b79a76c86db17e4e355:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7c028b79a76c86db17e4e355:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9a1e13e48c5f3401995d4aea:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9a1e13e48c5f3401995d4aea:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9a1e13e48c5f3401995d4aea:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9a1e13e48c5f3401995d4aea:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f39435f416b3a0bae23647a6:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f39435f416b3a0bae23647a6:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f39435f416b3a0bae23647a6:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f39435f416b3a0bae23647a6:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e85918d93b8f26c1d05dfe14:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e85918d93b8f26c1d05dfe14:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e85918d93b8f26c1d05dfe14:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e85918d93b8f26c1d05dfe14:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 224, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 221, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9421d90164ee0257fe3e727d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9421d90164ee0257fe3e727d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9421d90164ee0257fe3e727d:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9421d90164ee0257fe3e727d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d8dda02461c89b190e12b0c9:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d8dda02461c89b190e12b0c9:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d8dda02461c89b190e12b0c9:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d8dda02461c89b190e12b0c9:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:64d53b0831fecd4af36767e0:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:64d53b0831fecd4af36767e0:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:64d53b0831fecd4af36767e0:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:64d53b0831fecd4af36767e0:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3664759b6633c11ca5d8e36d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3664759b6633c11ca5d8e36d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3664759b6633c11ca5d8e36d:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3664759b6633c11ca5d8e36d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 189, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 195, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 208, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 189, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 195, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 209, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 189, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 195, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 207, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 188, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 194, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5e7387b379d5ac38a3303d86:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5e7387b379d5ac38a3303d86:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5e7387b379d5ac38a3303d86:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5e7387b379d5ac38a3303d86:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:91c593992ed356c02167928f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:91c593992ed356c02167928f:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:91c593992ed356c02167928f:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:91c593992ed356c02167928f:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b884ef2e8f42ac2d63d357e2:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b884ef2e8f42ac2d63d357e2:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b884ef2e8f42ac2d63d357e2:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b884ef2e8f42ac2d63d357e2:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7deaeb6a65c92f55a9839aee:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7deaeb6a65c92f55a9839aee:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7deaeb6a65c92f55a9839aee:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7deaeb6a65c92f55a9839aee:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 221, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 236, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e19f21eb60e12080c813e7d5:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e19f21eb60e12080c813e7d5:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e19f21eb60e12080c813e7d5:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e19f21eb60e12080c813e7d5:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b588a89600945856a65a65d5:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b588a89600945856a65a65d5:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b588a89600945856a65a65d5:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b588a89600945856a65a65d5:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:807ed6b55bdbd5cae8404a44:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:807ed6b55bdbd5cae8404a44:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:807ed6b55bdbd5cae8404a44:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:807ed6b55bdbd5cae8404a44:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:849d3f6680769a349ea6804e:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:849d3f6680769a349ea6804e:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:849d3f6680769a349ea6804e:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:849d3f6680769a349ea6804e:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 228, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 231, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 228, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:96b29694b8f8be77c05a7ad4:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:96b29694b8f8be77c05a7ad4:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:96b29694b8f8be77c05a7ad4:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:96b29694b8f8be77c05a7ad4:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ab770b0e55dc3a2c83839230:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ab770b0e55dc3a2c83839230:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ab770b0e55dc3a2c83839230:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ab770b0e55dc3a2c83839230:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e208c82a1a53aaa466a7d90b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e208c82a1a53aaa466a7d90b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e208c82a1a53aaa466a7d90b:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e208c82a1a53aaa466a7d90b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:289128d773eb9f635717f7df:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:289128d773eb9f635717f7df:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:289128d773eb9f635717f7df:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:289128d773eb9f635717f7df:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 225, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 226, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 239, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 237, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:158c1ec585e67397a4124aef:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:158c1ec585e67397a4124aef:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:158c1ec585e67397a4124aef:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:158c1ec585e67397a4124aef:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3b9af3823f55bc1710f3ffbf:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3b9af3823f55bc1710f3ffbf:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3b9af3823f55bc1710f3ffbf:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3b9af3823f55bc1710f3ffbf:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e7dc5b5e4133d3518dd65a04:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e7dc5b5e4133d3518dd65a04:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e7dc5b5e4133d3518dd65a04:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e7dc5b5e4133d3518dd65a04:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c2ed8fcd51b7007b0c3d07b5:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c2ed8fcd51b7007b0c3d07b5:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c2ed8fcd51b7007b0c3d07b5:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c2ed8fcd51b7007b0c3d07b5:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 218, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 221, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 200, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 217, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:28659e15e2e3bd051af50b5b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:28659e15e2e3bd051af50b5b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:28659e15e2e3bd051af50b5b:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:28659e15e2e3bd051af50b5b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:64cf7df1eefe1400e72f81cb:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:64cf7df1eefe1400e72f81cb:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:64cf7df1eefe1400e72f81cb:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:64cf7df1eefe1400e72f81cb:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:83a83973fd86ddcd08278c77:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:83a83973fd86ddcd08278c77:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:83a83973fd86ddcd08278c77:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:83a83973fd86ddcd08278c77:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e8e18f55d8092c0ee71b36a4:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e8e18f55d8092c0ee71b36a4:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e8e18f55d8092c0ee71b36a4:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e8e18f55d8092c0ee71b36a4:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 559, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 559, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 559, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 559, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 546, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 546, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 560, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 547, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 561, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:587e940982bc37d3a9d3ebbc:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:587e940982bc37d3a9d3ebbc:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:587e940982bc37d3a9d3ebbc:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:587e940982bc37d3a9d3ebbc:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e5222ca460d01b03e75a1acf:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e5222ca460d01b03e75a1acf:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e5222ca460d01b03e75a1acf:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e5222ca460d01b03e75a1acf:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:54224971d3a7ebdd73c88f5f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:54224971d3a7ebdd73c88f5f:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:54224971d3a7ebdd73c88f5f:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:54224971d3a7ebdd73c88f5f:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:548c1d7ff2b216307d82118d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:548c1d7ff2b216307d82118d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:548c1d7ff2b216307d82118d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:548c1d7ff2b216307d82118d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 475, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 481, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 496, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 476, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 482, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 498, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 477, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 483, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 474, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 480, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f602fc8e15f8718804a47a1a:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f602fc8e15f8718804a47a1a:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f602fc8e15f8718804a47a1a:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f602fc8e15f8718804a47a1a:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4945a0bf9e0f77491e12fa68:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4945a0bf9e0f77491e12fa68:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4945a0bf9e0f77491e12fa68:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4945a0bf9e0f77491e12fa68:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3c58cce3a6bc409ab2407674:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3c58cce3a6bc409ab2407674:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3c58cce3a6bc409ab2407674:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3c58cce3a6bc409ab2407674:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1fa03fee2824f53457cd8e6d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1fa03fee2824f53457cd8e6d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1fa03fee2824f53457cd8e6d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1fa03fee2824f53457cd8e6d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 522, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 502, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 523, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 502, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 536, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 521, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 501, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 535, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 535, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 535, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 535, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:21bd326ce12d821777866873:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:21bd326ce12d821777866873:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:21bd326ce12d821777866873:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:21bd326ce12d821777866873:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5ae346163f4ccb67ad5df4cf:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5ae346163f4ccb67ad5df4cf:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5ae346163f4ccb67ad5df4cf:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5ae346163f4ccb67ad5df4cf:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9566442c05492d7ee04c89a0:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 186, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9566442c05492d7ee04c89a0:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9566442c05492d7ee04c89a0:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9566442c05492d7ee04c89a0:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:63673902a990ea078c840d71:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 188, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:63673902a990ea078c840d71:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:63673902a990ea078c840d71:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:63673902a990ea078c840d71:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 500, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 506, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 524, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 509, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 525, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 504, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 538, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 538, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 538, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 538, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:860d0c0789d6afbac0ce3bba:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:860d0c0789d6afbac0ce3bba:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:860d0c0789d6afbac0ce3bba:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:860d0c0789d6afbac0ce3bba:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:775256c2c671ac2e72ebdd24:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:775256c2c671ac2e72ebdd24:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:775256c2c671ac2e72ebdd24:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:775256c2c671ac2e72ebdd24:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:38e40a2c70bb7ec9751a2cdf:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:38e40a2c70bb7ec9751a2cdf:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:38e40a2c70bb7ec9751a2cdf:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:38e40a2c70bb7ec9751a2cdf:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a07bcdf7cf6871e5f690963:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a07bcdf7cf6871e5f690963:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a07bcdf7cf6871e5f690963:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a07bcdf7cf6871e5f690963:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 402, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 408, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 420, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 400, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 406, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 434, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 434, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 434, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 434, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 421, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 401, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 407, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 435, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 435, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 435, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 435, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 423, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 402, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 408, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6a67db98c3bc114af9f92771:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6a67db98c3bc114af9f92771:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6a67db98c3bc114af9f92771:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6a67db98c3bc114af9f92771:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:14fc341d4676da4ba0c2c856:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:14fc341d4676da4ba0c2c856:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:14fc341d4676da4ba0c2c856:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:14fc341d4676da4ba0c2c856:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:71ddc8139db3932dc53395ea:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:71ddc8139db3932dc53395ea:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:71ddc8139db3932dc53395ea:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:71ddc8139db3932dc53395ea:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:eeeb37680018cbca5c0ce537:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:eeeb37680018cbca5c0ce537:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:eeeb37680018cbca5c0ce537:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:eeeb37680018cbca5c0ce537:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 505, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 505, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 517, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 498, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 504, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 500, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 506, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:16e173e44fd1eb833e530644:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:16e173e44fd1eb833e530644:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:16e173e44fd1eb833e530644:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:16e173e44fd1eb833e530644:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ee03d4da232d5808058554c5:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ee03d4da232d5808058554c5:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ee03d4da232d5808058554c5:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ee03d4da232d5808058554c5:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d09b8ff53ace600b84d8a0ad:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d09b8ff53ace600b84d8a0ad:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d09b8ff53ace600b84d8a0ad:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d09b8ff53ace600b84d8a0ad:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6e78fa8aa3855cf9e9a8b2f6:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6e78fa8aa3855cf9e9a8b2f6:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6e78fa8aa3855cf9e9a8b2f6:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6e78fa8aa3855cf9e9a8b2f6:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 517, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 502, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 514, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 495, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 501, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 518, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 518, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a336313d612f3aa9c3be807b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a336313d612f3aa9c3be807b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a336313d612f3aa9c3be807b:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a336313d612f3aa9c3be807b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0c81cbbe176f913d0b8e93ea:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0c81cbbe176f913d0b8e93ea:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0c81cbbe176f913d0b8e93ea:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0c81cbbe176f913d0b8e93ea:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6ada38c77274320ad626e980:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6ada38c77274320ad626e980:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6ada38c77274320ad626e980:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6ada38c77274320ad626e980:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:403dd328ac9cdb6fc3a2d912:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:403dd328ac9cdb6fc3a2d912:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:403dd328ac9cdb6fc3a2d912:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:403dd328ac9cdb6fc3a2d912:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 491, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 512, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 498, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 512, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 498, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 526, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 510, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:870e7a8c20e72999a38f1bbb:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:870e7a8c20e72999a38f1bbb:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:870e7a8c20e72999a38f1bbb:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:870e7a8c20e72999a38f1bbb:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:761db75641266f665d423dd6:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:761db75641266f665d423dd6:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:761db75641266f665d423dd6:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:761db75641266f665d423dd6:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f3d23931d79320f7dcb6b809:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f3d23931d79320f7dcb6b809:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f3d23931d79320f7dcb6b809:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f3d23931d79320f7dcb6b809:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:733f1c4c8d36eef7b6a3a628:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:733f1c4c8d36eef7b6a3a628:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:733f1c4c8d36eef7b6a3a628:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:733f1c4c8d36eef7b6a3a628:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 399, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 405, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 419, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 399, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 405, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 420, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 399, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 405, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 433, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 418, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 398, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 404, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 432, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 432, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 432, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 432, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58998e2b30200567009458f2:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58998e2b30200567009458f2:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58998e2b30200567009458f2:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58998e2b30200567009458f2:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d20444edd7bce4c96411c22d:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d20444edd7bce4c96411c22d:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d20444edd7bce4c96411c22d:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d20444edd7bce4c96411c22d:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e9f2b3223460fa3d568ffc9f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e9f2b3223460fa3d568ffc9f:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e9f2b3223460fa3d568ffc9f:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e9f2b3223460fa3d568ffc9f:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:621b671d54f22d05c2855b2b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:621b671d54f22d05c2855b2b:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:621b671d54f22d05c2855b2b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:621b671d54f22d05c2855b2b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 487, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 521, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 521, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 521, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 521, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 507, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 506, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a90f25174a8a6744aac7d061:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a90f25174a8a6744aac7d061:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a90f25174a8a6744aac7d061:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a90f25174a8a6744aac7d061:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3a60ced263839ebe4ffa236f:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3a60ced263839ebe4ffa236f:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3a60ced263839ebe4ffa236f:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3a60ced263839ebe4ffa236f:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5566ec4180230f4b928c1c51:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5566ec4180230f4b928c1c51:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5566ec4180230f4b928c1c51:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5566ec4180230f4b928c1c51:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:17d6eed389e537292d04e135:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 187, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:17d6eed389e537292d04e135:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:17d6eed389e537292d04e135:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:17d6eed389e537292d04e135:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 516, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 498, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 504, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 518, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 498, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 504, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 532, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 517, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 503, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 531, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e6fc77aaa094d0d78957065e:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e6fc77aaa094d0d78957065e:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e6fc77aaa094d0d78957065e:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e6fc77aaa094d0d78957065e:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7b1c6353934e5e0b84835f97:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7b1c6353934e5e0b84835f97:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7b1c6353934e5e0b84835f97:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7b1c6353934e5e0b84835f97:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:69fefba3afa3e52436b6a3d9:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:69fefba3afa3e52436b6a3d9:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:69fefba3afa3e52436b6a3d9:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:69fefba3afa3e52436b6a3d9:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 476, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 482, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 497, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 477, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 483, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 496, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 477, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 483, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:action", + "split": "train", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 495, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:solvable", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 476, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:value", + "split": "train", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 482, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:11f37c8a34e52be1e5290999:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:11f37c8a34e52be1e5290999:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:11f37c8a34e52be1e5290999:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:11f37c8a34e52be1e5290999:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fda49c37ac5bc4d0ace320fa:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fda49c37ac5bc4d0ace320fa:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fda49c37ac5bc4d0ace320fa:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fda49c37ac5bc4d0ace320fa:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:389deeada306290bf795e0c1:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:389deeada306290bf795e0c1:safe_north", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:389deeada306290bf795e0c1:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:389deeada306290bf795e0c1:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:01ae3faf9d37bf799bde6b7b:action", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:01ae3faf9d37bf799bde6b7b:safe_east", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:01ae3faf9d37bf799bde6b7b:safe_south", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:01ae3faf9d37bf799bde6b7b:safe_west", + "split": "train", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a8b7df91259c3f05aeb382b4:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a8b7df91259c3f05aeb382b4:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a8b7df91259c3f05aeb382b4:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a8b7df91259c3f05aeb382b4:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:df90ae5ad4e7c29cddbeece3:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:df90ae5ad4e7c29cddbeece3:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:df90ae5ad4e7c29cddbeece3:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:df90ae5ad4e7c29cddbeece3:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:cd94f508a6fac43a109bbad1:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:cd94f508a6fac43a109bbad1:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:cd94f508a6fac43a109bbad1:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:cd94f508a6fac43a109bbad1:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 150, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 131, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 165, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c388fa806f00892833d237ff:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c388fa806f00892833d237ff:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c388fa806f00892833d237ff:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c388fa806f00892833d237ff:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4c5c5ed5f1c363a3abb500d2:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4c5c5ed5f1c363a3abb500d2:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4c5c5ed5f1c363a3abb500d2:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4c5c5ed5f1c363a3abb500d2:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:576e0861df9167b47f30ee2e:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:576e0861df9167b47f30ee2e:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:576e0861df9167b47f30ee2e:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:576e0861df9167b47f30ee2e:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:52fbbba9d2809da41935b256:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:52fbbba9d2809da41935b256:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:52fbbba9d2809da41935b256:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:52fbbba9d2809da41935b256:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:67cc2d1e520f345a77c25258:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:67cc2d1e520f345a77c25258:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:67cc2d1e520f345a77c25258:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:67cc2d1e520f345a77c25258:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4e0dd4fb1004b1b5f10505fb:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4e0dd4fb1004b1b5f10505fb:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4e0dd4fb1004b1b5f10505fb:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4e0dd4fb1004b1b5f10505fb:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a156b999ce6a51e078e7d63b:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a156b999ce6a51e078e7d63b:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a156b999ce6a51e078e7d63b:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a156b999ce6a51e078e7d63b:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9471f04d7505354571f2afae:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9471f04d7505354571f2afae:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9471f04d7505354571f2afae:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9471f04d7505354571f2afae:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:78b599638feb2b5738052f42:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:78b599638feb2b5738052f42:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:78b599638feb2b5738052f42:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:78b599638feb2b5738052f42:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d2e3d53861d338d18932e5c3:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d2e3d53861d338d18932e5c3:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d2e3d53861d338d18932e5c3:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d2e3d53861d338d18932e5c3:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:655d2c7cb308d413d245f479:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:655d2c7cb308d413d245f479:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:655d2c7cb308d413d245f479:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:655d2c7cb308d413d245f479:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e1714aca03074946552e86d7:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e1714aca03074946552e86d7:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e1714aca03074946552e86d7:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e1714aca03074946552e86d7:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 188, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 194, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 190, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 196, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 224, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 224, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 224, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 224, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 208, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 188, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 194, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 189, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 195, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7f5e2225e29d3de69626ddb4:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7f5e2225e29d3de69626ddb4:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7f5e2225e29d3de69626ddb4:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7f5e2225e29d3de69626ddb4:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ca33cb36c8613e8697afb773:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ca33cb36c8613e8697afb773:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ca33cb36c8613e8697afb773:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ca33cb36c8613e8697afb773:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b7664a71002acbd88104c520:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b7664a71002acbd88104c520:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b7664a71002acbd88104c520:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b7664a71002acbd88104c520:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:57d32914072132dfcd28dde9:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:57d32914072132dfcd28dde9:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:57d32914072132dfcd28dde9:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:57d32914072132dfcd28dde9:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 228, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 229, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 243, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c8567147711d9e316a6aa4cf:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c8567147711d9e316a6aa4cf:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c8567147711d9e316a6aa4cf:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c8567147711d9e316a6aa4cf:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e09d6920ac6a46a73bb0a67:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 177, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e09d6920ac6a46a73bb0a67:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e09d6920ac6a46a73bb0a67:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e09d6920ac6a46a73bb0a67:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:432bf9e1b22aee6ca9a73089:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:432bf9e1b22aee6ca9a73089:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:432bf9e1b22aee6ca9a73089:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:432bf9e1b22aee6ca9a73089:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6d101d00264544fd0e797201:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6d101d00264544fd0e797201:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6d101d00264544fd0e797201:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6d101d00264544fd0e797201:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 196, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 196, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 218, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 198, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 232, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 217, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 197, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0832e3f7e20036db9aa11da9:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0832e3f7e20036db9aa11da9:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0832e3f7e20036db9aa11da9:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0832e3f7e20036db9aa11da9:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:994f831336c3c300105f9539:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:994f831336c3c300105f9539:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:994f831336c3c300105f9539:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:994f831336c3c300105f9539:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:77649ea6ad51f9563a36480c:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:77649ea6ad51f9563a36480c:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:77649ea6ad51f9563a36480c:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:77649ea6ad51f9563a36480c:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:148cdef4390f40347ee09888:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:148cdef4390f40347ee09888:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:148cdef4390f40347ee09888:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:148cdef4390f40347ee09888:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 227, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 229, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 257, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 257, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 257, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 257, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 242, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 228, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 256, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 256, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 256, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 256, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 240, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 227, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 255, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8f71ea5f3a25834336bcc05d:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8f71ea5f3a25834336bcc05d:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8f71ea5f3a25834336bcc05d:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8f71ea5f3a25834336bcc05d:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6ac6bd536b0757091c5455f0:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6ac6bd536b0757091c5455f0:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6ac6bd536b0757091c5455f0:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6ac6bd536b0757091c5455f0:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:964476c58e1e5730badc9744:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:964476c58e1e5730badc9744:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:964476c58e1e5730badc9744:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:964476c58e1e5730badc9744:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7a57be18ed7438683c73015f:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7a57be18ed7438683c73015f:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7a57be18ed7438683c73015f:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7a57be18ed7438683c73015f:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 423, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 434, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 415, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 421, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 449, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 449, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 449, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 449, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 438, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 417, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 423, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 451, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 436, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 416, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 422, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 450, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 450, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 450, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 450, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:051e6e8a2266fa53121d8477:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:051e6e8a2266fa53121d8477:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:051e6e8a2266fa53121d8477:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:051e6e8a2266fa53121d8477:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0f28d07901bcbe13b678ee02:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0f28d07901bcbe13b678ee02:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0f28d07901bcbe13b678ee02:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0f28d07901bcbe13b678ee02:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e7ba57b10725beea1c1e6143:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e7ba57b10725beea1c1e6143:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e7ba57b10725beea1c1e6143:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e7ba57b10725beea1c1e6143:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3d0675a33c07627aadd5584d:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3d0675a33c07627aadd5584d:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3d0675a33c07627aadd5584d:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3d0675a33c07627aadd5584d:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 513, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 513, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 541, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 529, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 514, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 514, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 542, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:883677764232a4c9ad39f564:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:883677764232a4c9ad39f564:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:883677764232a4c9ad39f564:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:883677764232a4c9ad39f564:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4ce2ba2de3d075d05fbf6931:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4ce2ba2de3d075d05fbf6931:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4ce2ba2de3d075d05fbf6931:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4ce2ba2de3d075d05fbf6931:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:49bd5f26505a7ebdca09e2e6:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:49bd5f26505a7ebdca09e2e6:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:49bd5f26505a7ebdca09e2e6:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:49bd5f26505a7ebdca09e2e6:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8bb9604a0be751b4ad25ceae:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8bb9604a0be751b4ad25ceae:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8bb9604a0be751b4ad25ceae:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8bb9604a0be751b4ad25ceae:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 500, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 515, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 495, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 501, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 529, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 517, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 502, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 530, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 514, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 500, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c0df79c78d46b682825412f6:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c0df79c78d46b682825412f6:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c0df79c78d46b682825412f6:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c0df79c78d46b682825412f6:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:11d45c3c68b9900a8ac18e46:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:11d45c3c68b9900a8ac18e46:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:11d45c3c68b9900a8ac18e46:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:11d45c3c68b9900a8ac18e46:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:085a2608e9718db6826e4fcd:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:085a2608e9718db6826e4fcd:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:085a2608e9718db6826e4fcd:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:085a2608e9718db6826e4fcd:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 344, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 350, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 367, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 346, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 352, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 380, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 380, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 380, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 380, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 364, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 344, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 350, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 378, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:action", + "split": "dev", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 368, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:solvable", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 347, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:value", + "split": "dev", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 353, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 381, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 381, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 381, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 381, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:092d50b97340fa17002cb4ab:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:092d50b97340fa17002cb4ab:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:092d50b97340fa17002cb4ab:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:092d50b97340fa17002cb4ab:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:417fceca334f03fa9927ab18:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:417fceca334f03fa9927ab18:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:417fceca334f03fa9927ab18:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:417fceca334f03fa9927ab18:safe_west", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:55c505682323e7ba8a0a15c0:action", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:55c505682323e7ba8a0a15c0:safe_north", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:55c505682323e7ba8a0a15c0:safe_east", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:55c505682323e7ba8a0a15c0:safe_south", + "split": "dev", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:75a4f70d771aa4b61f9c0d3b:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:75a4f70d771aa4b61f9c0d3b:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:75a4f70d771aa4b61f9c0d3b:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:75a4f70d771aa4b61f9c0d3b:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:01ff75816b735bcdd1aeda73:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:01ff75816b735bcdd1aeda73:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:01ff75816b735bcdd1aeda73:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:01ff75816b735bcdd1aeda73:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9261d348d2b3d6d8b0a34f81:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9261d348d2b3d6d8b0a34f81:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9261d348d2b3d6d8b0a34f81:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9261d348d2b3d6d8b0a34f81:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d8903c26d0c9b4585993dd4f:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 188, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d8903c26d0c9b4585993dd4f:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d8903c26d0c9b4585993dd4f:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d8903c26d0c9b4585993dd4f:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 151, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 132, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 138, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 166, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0cde1d49787ade977529f5d1:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0cde1d49787ade977529f5d1:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0cde1d49787ade977529f5d1:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0cde1d49787ade977529f5d1:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58a227e097f6ef3b33076919:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58a227e097f6ef3b33076919:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58a227e097f6ef3b33076919:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58a227e097f6ef3b33076919:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:08b1f28ba1a403013512cf46:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:08b1f28ba1a403013512cf46:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:08b1f28ba1a403013512cf46:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:08b1f28ba1a403013512cf46:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:abb1527dca73ee5804fa14e9:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:abb1527dca73ee5804fa14e9:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:abb1527dca73ee5804fa14e9:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:abb1527dca73ee5804fa14e9:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:070695834c3b294ad333f6ec:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:070695834c3b294ad333f6ec:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:070695834c3b294ad333f6ec:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:070695834c3b294ad333f6ec:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f8db42503df5695a2a3ed03d:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f8db42503df5695a2a3ed03d:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f8db42503df5695a2a3ed03d:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f8db42503df5695a2a3ed03d:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0a57c6db45c1f166083e73d8:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0a57c6db45c1f166083e73d8:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0a57c6db45c1f166083e73d8:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0a57c6db45c1f166083e73d8:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:63c079a7af785193d2dcb0a2:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:63c079a7af785193d2dcb0a2:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:63c079a7af785193d2dcb0a2:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:63c079a7af785193d2dcb0a2:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e899f3d6ecef3a09d8ffc119:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e899f3d6ecef3a09d8ffc119:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e899f3d6ecef3a09d8ffc119:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e899f3d6ecef3a09d8ffc119:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 197, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 197, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 203, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 215, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 196, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 196, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:186c9eeb1aae95c0b0c3ef44:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:186c9eeb1aae95c0b0c3ef44:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:186c9eeb1aae95c0b0c3ef44:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:186c9eeb1aae95c0b0c3ef44:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ce254f8faea7d887f1a666f3:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ce254f8faea7d887f1a666f3:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ce254f8faea7d887f1a666f3:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ce254f8faea7d887f1a666f3:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:09e3bf5fdec487c1c6f5af08:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 187, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:09e3bf5fdec487c1c6f5af08:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:09e3bf5fdec487c1c6f5af08:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:09e3bf5fdec487c1c6f5af08:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 231, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 230, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 245, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 229, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:51b44ad4a37cc5094c2900be:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:51b44ad4a37cc5094c2900be:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:51b44ad4a37cc5094c2900be:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:51b44ad4a37cc5094c2900be:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b884811f4367bae3e1ebdd7a:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b884811f4367bae3e1ebdd7a:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b884811f4367bae3e1ebdd7a:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b884811f4367bae3e1ebdd7a:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dc81c3fab4d17b908531d011:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dc81c3fab4d17b908531d011:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dc81c3fab4d17b908531d011:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dc81c3fab4d17b908531d011:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a8e9c88494add4eff7bdca3:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a8e9c88494add4eff7bdca3:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a8e9c88494add4eff7bdca3:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a8e9c88494add4eff7bdca3:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 224, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 223, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 225, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 238, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e6867817864cb9193a8ebf1:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e6867817864cb9193a8ebf1:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e6867817864cb9193a8ebf1:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e6867817864cb9193a8ebf1:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:22cc42ee4a862f50bec2e5b5:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:22cc42ee4a862f50bec2e5b5:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:22cc42ee4a862f50bec2e5b5:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:22cc42ee4a862f50bec2e5b5:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dbf416426f207e29ea1c55cf:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dbf416426f207e29ea1c55cf:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dbf416426f207e29ea1c55cf:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:dbf416426f207e29ea1c55cf:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e0c7592c01c85392eb64bb73:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e0c7592c01c85392eb64bb73:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e0c7592c01c85392eb64bb73:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e0c7592c01c85392eb64bb73:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 228, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 227, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 242, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 227, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 241, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a00fe6dca724d788d40bbba7:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a00fe6dca724d788d40bbba7:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a00fe6dca724d788d40bbba7:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a00fe6dca724d788d40bbba7:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:04cbd162811373e57f53152b:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:04cbd162811373e57f53152b:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:04cbd162811373e57f53152b:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:04cbd162811373e57f53152b:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:309dbe3ed47e8d1c5a17d841:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:309dbe3ed47e8d1c5a17d841:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:309dbe3ed47e8d1c5a17d841:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:309dbe3ed47e8d1c5a17d841:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5dd7b1aef1af8e9ff190595f:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5dd7b1aef1af8e9ff190595f:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5dd7b1aef1af8e9ff190595f:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5dd7b1aef1af8e9ff190595f:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 491, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 514, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 512, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 491, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:67d7d8cd30a917dd7060d049:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:67d7d8cd30a917dd7060d049:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:67d7d8cd30a917dd7060d049:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:67d7d8cd30a917dd7060d049:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:850d0c734d3afba2f5d77144:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:850d0c734d3afba2f5d77144:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:850d0c734d3afba2f5d77144:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:850d0c734d3afba2f5d77144:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0d69e28ac9cee01ce099a94c:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0d69e28ac9cee01ce099a94c:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0d69e28ac9cee01ce099a94c:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0d69e28ac9cee01ce099a94c:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:df1076fd97ec6409895096ad:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:df1076fd97ec6409895096ad:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:df1076fd97ec6409895096ad:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:df1076fd97ec6409895096ad:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 510, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 511, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 496, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:09d01d5cb2826ee7d7b75692:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:09d01d5cb2826ee7d7b75692:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:09d01d5cb2826ee7d7b75692:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:09d01d5cb2826ee7d7b75692:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6522cd24159d9f67a110fb91:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6522cd24159d9f67a110fb91:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6522cd24159d9f67a110fb91:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6522cd24159d9f67a110fb91:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8a431ce79c63ab176847b014:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8a431ce79c63ab176847b014:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8a431ce79c63ab176847b014:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8a431ce79c63ab176847b014:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5fef7acaca5a9f9ea2b6dce3:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5fef7acaca5a9f9ea2b6dce3:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5fef7acaca5a9f9ea2b6dce3:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:5fef7acaca5a9f9ea2b6dce3:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 484, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 490, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 518, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 518, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 518, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 518, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 506, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 486, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 492, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 520, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 505, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 485, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 491, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:92c385cd4eb8c43cda20267f:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:92c385cd4eb8c43cda20267f:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:92c385cd4eb8c43cda20267f:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:92c385cd4eb8c43cda20267f:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1c07f3a9c4269d3d224fa63b:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1c07f3a9c4269d3d224fa63b:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1c07f3a9c4269d3d224fa63b:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1c07f3a9c4269d3d224fa63b:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6bac4198ff50afdd1c28bf82:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 186, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6bac4198ff50afdd1c28bf82:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6bac4198ff50afdd1c28bf82:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6bac4198ff50afdd1c28bf82:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4d5d9a736a2c010ecddf3bd7:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4d5d9a736a2c010ecddf3bd7:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4d5d9a736a2c010ecddf3bd7:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4d5d9a736a2c010ecddf3bd7:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 505, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 520, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 500, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 506, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 534, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 519, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 505, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:action", + "split": "calibration", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 518, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:solvable", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:value", + "split": "calibration", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 505, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 533, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ec8cd8b3711ab660e4121bfd:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ec8cd8b3711ab660e4121bfd:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ec8cd8b3711ab660e4121bfd:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ec8cd8b3711ab660e4121bfd:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:611b5d26ce10b988029a6daf:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:611b5d26ce10b988029a6daf:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:611b5d26ce10b988029a6daf:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:611b5d26ce10b988029a6daf:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a408545a8ce348a3053aa8ed:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a408545a8ce348a3053aa8ed:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a408545a8ce348a3053aa8ed:safe_east", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a408545a8ce348a3053aa8ed:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9828eefaea5a58f642317f6e:action", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9828eefaea5a58f642317f6e:safe_north", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9828eefaea5a58f642317f6e:safe_south", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9828eefaea5a58f642317f6e:safe_west", + "split": "calibration", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0767d5e59aed5fd05cac2e9f:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0767d5e59aed5fd05cac2e9f:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0767d5e59aed5fd05cac2e9f:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0767d5e59aed5fd05cac2e9f:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4f81c7dda2b7d32567d91775:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4f81c7dda2b7d32567d91775:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4f81c7dda2b7d32567d91775:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4f81c7dda2b7d32567d91775:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:852918c7ae73c068c0ff56c9:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:852918c7ae73c068c0ff56c9:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:852918c7ae73c068c0ff56c9:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:852918c7ae73c068c0ff56c9:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:635d2abfc307ec0710e44f69:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:635d2abfc307ec0710e44f69:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:635d2abfc307ec0710e44f69:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:635d2abfc307ec0710e44f69:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 128, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 148, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 129, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 163, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 147, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 128, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 147, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 128, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 162, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6b804b8015033d9a92a4dec1:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6b804b8015033d9a92a4dec1:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6b804b8015033d9a92a4dec1:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6b804b8015033d9a92a4dec1:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c36d051be7af4b360d34b50d:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c36d051be7af4b360d34b50d:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c36d051be7af4b360d34b50d:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c36d051be7af4b360d34b50d:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8cac327e96392079d7260378:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8cac327e96392079d7260378:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8cac327e96392079d7260378:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:8cac327e96392079d7260378:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0d2c2cfdf6e63083ab122e25:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0d2c2cfdf6e63083ab122e25:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0d2c2cfdf6e63083ab122e25:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:0d2c2cfdf6e63083ab122e25:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 136, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 142, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 170, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 156, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 137, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 143, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 154, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 135, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 141, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 169, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:de8b76d31b3502f06ab7bece:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:de8b76d31b3502f06ab7bece:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:de8b76d31b3502f06ab7bece:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:de8b76d31b3502f06ab7bece:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b6f0bcae75ff0eedc4104dd0:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b6f0bcae75ff0eedc4104dd0:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b6f0bcae75ff0eedc4104dd0:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:b6f0bcae75ff0eedc4104dd0:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:2340031ad35a3132a389f44d:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:2340031ad35a3132a389f44d:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:2340031ad35a3132a389f44d:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:2340031ad35a3132a389f44d:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d76b23c4185e25ab14dda3c8:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 176, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d76b23c4185e25ab14dda3c8:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d76b23c4185e25ab14dda3c8:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d76b23c4185e25ab14dda3c8:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 152, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 133, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 139, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 167, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 153, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 134, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 140, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 168, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a9a97a95a878d441fd0eed1d:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 172, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a9a97a95a878d441fd0eed1d:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a9a97a95a878d441fd0eed1d:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a9a97a95a878d441fd0eed1d:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 202, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f4916f37b65f9a17a6e6d10b:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 180, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f4916f37b65f9a17a6e6d10b:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f4916f37b65f9a17a6e6d10b:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f4916f37b65f9a17a6e6d10b:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1f964939ebb08f4325841e5b:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1f964939ebb08f4325841e5b:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1f964939ebb08f4325841e5b:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1f964939ebb08f4325841e5b:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58485cd9064d09b8277665a0:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58485cd9064d09b8277665a0:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58485cd9064d09b8277665a0:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:58485cd9064d09b8277665a0:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 226, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 226, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 240, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 226, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 240, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 220, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 226, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 254, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e71d44addec992f5c8c8e8f0:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e71d44addec992f5c8c8e8f0:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e71d44addec992f5c8c8e8f0:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e71d44addec992f5c8c8e8f0:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d50907af6656e0c98fbf31da:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d50907af6656e0c98fbf31da:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d50907af6656e0c98fbf31da:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d50907af6656e0c98fbf31da:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e97eaa71c50f430fe64471d5:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e97eaa71c50f430fe64471d5:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e97eaa71c50f430fe64471d5:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e97eaa71c50f430fe64471d5:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 230, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 229, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 244, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 219, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 247, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 247, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 247, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 247, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f2f72dad5dedbdb36adc45a1:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f2f72dad5dedbdb36adc45a1:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f2f72dad5dedbdb36adc45a1:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f2f72dad5dedbdb36adc45a1:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4908b0108026b8d03e2e30b2:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4908b0108026b8d03e2e30b2:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4908b0108026b8d03e2e30b2:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4908b0108026b8d03e2e30b2:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e21f7a6a1b5a3f6776d566b:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e21f7a6a1b5a3f6776d566b:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e21f7a6a1b5a3f6776d566b:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e21f7a6a1b5a3f6776d566b:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c60247f3751938fe95596441:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c60247f3751938fe95596441:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c60247f3751938fe95596441:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:c60247f3751938fe95596441:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 233, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 220, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 235, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 234, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 220, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 248, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e272f5b22e1649dc9a40f56f:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e272f5b22e1649dc9a40f56f:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e272f5b22e1649dc9a40f56f:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e272f5b22e1649dc9a40f56f:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1a75df6d584a4d37af00c4f9:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1a75df6d584a4d37af00c4f9:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1a75df6d584a4d37af00c4f9:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:1a75df6d584a4d37af00c4f9:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ba01f7e9755f2a8bc3f51ad5:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 189, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ba01f7e9755f2a8bc3f51ad5:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ba01f7e9755f2a8bc3f51ad5:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ba01f7e9755f2a8bc3f51ad5:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ad228b61d8212dc6c4cffdf1:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ad228b61d8212dc6c4cffdf1:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ad228b61d8212dc6c4cffdf1:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:ad228b61d8212dc6c4cffdf1:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 237, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 217, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 223, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 251, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 236, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 222, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 250, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 234, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 215, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 221, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 249, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:af8239c080fc2cb3fba98d2b:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:af8239c080fc2cb3fba98d2b:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:af8239c080fc2cb3fba98d2b:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:af8239c080fc2cb3fba98d2b:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9c812cad9f595a262b7d3e62:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9c812cad9f595a262b7d3e62:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9c812cad9f595a262b7d3e62:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9c812cad9f595a262b7d3e62:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:243bdce11167f2cb54cd23a6:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:243bdce11167f2cb54cd23a6:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:243bdce11167f2cb54cd23a6:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:243bdce11167f2cb54cd23a6:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e2c971a95e7b505a5e37f7d9:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 175, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e2c971a95e7b505a5e37f7d9:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e2c971a95e7b505a5e37f7d9:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e2c971a95e7b505a5e37f7d9:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 205, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 474, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 480, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 473, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 479, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 495, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 474, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 480, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 508, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 473, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 479, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 507, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e194fc071b1fde4b9e783c83:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e194fc071b1fde4b9e783c83:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e194fc071b1fde4b9e783c83:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:e194fc071b1fde4b9e783c83:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fcf919116a7028c2c3286bf4:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fcf919116a7028c2c3286bf4:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fcf919116a7028c2c3286bf4:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fcf919116a7028c2c3286bf4:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:99b3c3c8700d99220a1d030c:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 185, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:99b3c3c8700d99220a1d030c:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:99b3c3c8700d99220a1d030c:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:99b3c3c8700d99220a1d030c:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:17afc77a551ef63112fd0716:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:17afc77a551ef63112fd0716:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:17afc77a551ef63112fd0716:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:17afc77a551ef63112fd0716:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 518, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 524, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 552, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 552, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 552, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 552, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 537, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 517, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 523, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 551, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 551, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 551, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 551, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 539, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 540, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 519, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 553, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a0aee04affa084fcaeeaf52:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a0aee04affa084fcaeeaf52:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a0aee04affa084fcaeeaf52:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4a0aee04affa084fcaeeaf52:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3b5df35292fcfcfb1d2eeb27:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 181, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3b5df35292fcfcfb1d2eeb27:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3b5df35292fcfcfb1d2eeb27:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3b5df35292fcfcfb1d2eeb27:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 209, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6a6399d8e17bf0a9bc5100f2:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6a6399d8e17bf0a9bc5100f2:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6a6399d8e17bf0a9bc5100f2:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6a6399d8e17bf0a9bc5100f2:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3dea7e463701bc106a5b9e80:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3dea7e463701bc106a5b9e80:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3dea7e463701bc106a5b9e80:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3dea7e463701bc106a5b9e80:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 513, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 513, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 499, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 527, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 515, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 500, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 528, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 510, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 491, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 497, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 525, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9f940aae7bd55ce7e711ac49:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9f940aae7bd55ce7e711ac49:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9f940aae7bd55ce7e711ac49:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9f940aae7bd55ce7e711ac49:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:53c433339e8c5e57ea898a50:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 177, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:53c433339e8c5e57ea898a50:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:53c433339e8c5e57ea898a50:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:53c433339e8c5e57ea898a50:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 206, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e1043e67de35a1943feb328:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e1043e67de35a1943feb328:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e1043e67de35a1943feb328:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:9e1043e67de35a1943feb328:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:780daa690f231eb74dd97dca:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:780daa690f231eb74dd97dca:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:780daa690f231eb74dd97dca:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:780daa690f231eb74dd97dca:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 459, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 465, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 4, + "max_path_tokens": 480, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 459, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 465, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 493, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 480, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 460, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 466, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:action", + "split": "test", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 481, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:solvable", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 460, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:value", + "split": "test", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 466, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 494, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7a64e76f56a37f3684f5f79f:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7a64e76f56a37f3684f5f79f:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7a64e76f56a37f3684f5f79f:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7a64e76f56a37f3684f5f79f:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:70a97174b2c166c712894d0b:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:70a97174b2c166c712894d0b:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:70a97174b2c166c712894d0b:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:70a97174b2c166c712894d0b:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:57971912a78a6cac83176839:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 187, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:57971912a78a6cac83176839:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:57971912a78a6cac83176839:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:57971912a78a6cac83176839:safe_west", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 216, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fd0cb58bb748915f35c51f85:action", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fd0cb58bb748915f35c51f85:safe_north", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fd0cb58bb748915f35c51f85:safe_east", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fd0cb58bb748915f35c51f85:safe_south", + "split": "test", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 730, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 736, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 751, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 730, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 736, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 751, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 730, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 736, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 764, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 752, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 731, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 737, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 765, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 765, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 765, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 765, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3457a5bcfc7ae7e65820c327:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 182, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3457a5bcfc7ae7e65820c327:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3457a5bcfc7ae7e65820c327:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:3457a5bcfc7ae7e65820c327:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 210, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f2b50f863c17077d11d4476a:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 178, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f2b50f863c17077d11d4476a:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f2b50f863c17077d11d4476a:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:f2b50f863c17077d11d4476a:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 207, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:89685da0ba3d37f15f6a8d29:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 174, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:89685da0ba3d37f15f6a8d29:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:89685da0ba3d37f15f6a8d29:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:89685da0ba3d37f15f6a8d29:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 204, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a2aea55c8183660f23126df4:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 185, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a2aea55c8183660f23126df4:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a2aea55c8183660f23126df4:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a2aea55c8183660f23126df4:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1063, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1069, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 1085, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1064, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1070, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 1082, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1063, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1069, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1097, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 1085, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1064, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1070, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1098, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4323c20b90266bb4d9d253f9:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 183, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4323c20b90266bb4d9d253f9:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4323c20b90266bb4d9d253f9:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:4323c20b90266bb4d9d253f9:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 211, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:06405ef35d5208c85b1ebb1d:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 179, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:06405ef35d5208c85b1ebb1d:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:06405ef35d5208c85b1ebb1d:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:06405ef35d5208c85b1ebb1d:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 208, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a17f85e8992b92bfc18067aa:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 185, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a17f85e8992b92bfc18067aa:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a17f85e8992b92bfc18067aa:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a17f85e8992b92bfc18067aa:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 214, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d74471f2cd14288707392991:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d74471f2cd14288707392991:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d74471f2cd14288707392991:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:d74471f2cd14288707392991:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1053, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1059, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 1073, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1053, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1059, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1087, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 1074, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1054, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1060, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 1075, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1054, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1060, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1088, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:48047719c35c8036f2597b4c:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:48047719c35c8036f2597b4c:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:48047719c35c8036f2597b4c:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:48047719c35c8036f2597b4c:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6b7a640b5bdf1f583cf06e15:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6b7a640b5bdf1f583cf06e15:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6b7a640b5bdf1f583cf06e15:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:6b7a640b5bdf1f583cf06e15:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:20906b8454960dd9e5ef74c3:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:20906b8454960dd9e5ef74c3:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:20906b8454960dd9e5ef74c3:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:20906b8454960dd9e5ef74c3:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a2c757cd44b5e0873f48c513:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 190, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a2c757cd44b5e0873f48c513:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a2c757cd44b5e0873f48c513:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:a2c757cd44b5e0873f48c513:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 218, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 1027, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1007, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1013, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 1026, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1006, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1012, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 3, + "max_path_tokens": 1026, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1006, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1012, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1040, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:action", + "split": "ood", + "family_id": "scaled_maze", + "type": "choice", + "k": 2, + "max_path_tokens": 1028, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:solvable", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1007, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:value", + "split": "ood", + "family_id": "scaled_maze", + "type": "score", + "k": 7, + "max_path_tokens": 1013, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 1041, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7928813fe2cda7456d1a2d24:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7928813fe2cda7456d1a2d24:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7928813fe2cda7456d1a2d24:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:7928813fe2cda7456d1a2d24:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:56fae55f94b19332a280e632:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:56fae55f94b19332a280e632:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:56fae55f94b19332a280e632:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:56fae55f94b19332a280e632:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 212, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:910b45e72218ee95ef8b49f2:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:910b45e72218ee95ef8b49f2:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:910b45e72218ee95ef8b49f2:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:910b45e72218ee95ef8b49f2:safe_west", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fa0d0b06be9f8698a9622a2b:action", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "choice", + "k": 3, + "max_path_tokens": 184, + "gold_distribution_kind": "optimal_action_policy", + "gold_label_kind": "reference_argmax_compatibility", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fa0d0b06be9f8698a9622a2b:safe_north", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fa0d0b06be9f8698a9622a2b:safe_east", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "snake:fa0d0b06be9f8698a9622a2b:safe_south", + "split": "ood", + "family_id": "snake_one_step_safety_v4", + "type": "boolean", + "k": 2, + "max_path_tokens": 213, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + } +] diff --git a/variants/games_gold_seed17/tokenizer/chat_template.jinja b/variants/games_gold_seed17/tokenizer/chat_template.jinja new file mode 100644 index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4 --- /dev/null +++ b/variants/games_gold_seed17/tokenizer/chat_template.jinja @@ -0,0 +1,89 @@ +{%- if tools %} + {{- '<|im_start|>system\n' }} + {%- if messages[0].role == 'system' %} + {{- messages[0].content + '\n\n' }} + {%- endif %} + {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within XML tags:\n" }} + {%- for tool in tools %} + {{- "\n" }} + {{- tool | tojson }} + {%- endfor %} + {{- "\n\n\nFor each function call, return a json object with function name and arguments within XML tags:\n\n{\"name\": , \"arguments\": }\n<|im_end|>\n" }} +{%- else %} + {%- if messages[0].role == 'system' %} + {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }} + {%- endif %} +{%- endif %} +{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %} +{%- for message in messages[::-1] %} + {%- set index = (messages|length - 1) - loop.index0 %} + {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('') and message.content.endswith('')) %} + {%- set ns.multi_step_tool = false %} + {%- set ns.last_query_index = index %} + {%- endif %} +{%- endfor %} +{%- for message in messages %} + {%- if message.content is string %} + {%- set content = message.content %} + {%- else %} + {%- set content = '' %} + {%- endif %} + {%- if (message.role == "user") or (message.role == "system" and not loop.first) %} + {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }} + {%- elif message.role == "assistant" %} + {%- set reasoning_content = '' %} + {%- if message.reasoning_content is string %} + {%- set reasoning_content = message.reasoning_content %} + {%- else %} + {%- if '' in content %} + {%- set reasoning_content = content.split('')[0].rstrip('\n').split('')[-1].lstrip('\n') %} + {%- set content = content.split('')[-1].lstrip('\n') %} + {%- endif %} + {%- endif %} + {%- if loop.index0 > ns.last_query_index %} + {%- if loop.last or (not loop.last and reasoning_content) %} + {{- '<|im_start|>' + message.role + '\n\n' + reasoning_content.strip('\n') + '\n\n\n' + content.lstrip('\n') }} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- if message.tool_calls %} + {%- for tool_call in message.tool_calls %} + {%- if (loop.first and content) or (not loop.first) %} + {{- '\n' }} + {%- endif %} + {%- if tool_call.function %} + {%- set tool_call = tool_call.function %} + {%- endif %} + {{- '\n{"name": "' }} + {{- tool_call.name }} + {{- '", "arguments": ' }} + {%- if tool_call.arguments is string %} + {{- tool_call.arguments }} + {%- else %} + {{- tool_call.arguments | tojson }} + {%- endif %} + {{- '}\n' }} + {%- endfor %} + {%- endif %} + {{- '<|im_end|>\n' }} + {%- elif message.role == "tool" %} + {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %} + {{- '<|im_start|>user' }} + {%- endif %} + {{- '\n\n' }} + {{- content }} + {{- '\n' }} + {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %} + {{- '<|im_end|>\n' }} + {%- endif %} + {%- endif %} +{%- endfor %} +{%- if add_generation_prompt %} + {{- '<|im_start|>assistant\n' }} + {%- if enable_thinking is defined and enable_thinking is false %} + {{- '\n\n\n\n' }} + {%- endif %} +{%- endif %} \ No newline at end of file diff --git a/variants/games_gold_seed17/tokenizer/tokenizer.json b/variants/games_gold_seed17/tokenizer/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2 --- /dev/null +++ b/variants/games_gold_seed17/tokenizer/tokenizer.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506 +size 11422650 diff --git a/variants/games_gold_seed17/tokenizer/tokenizer_config.json b/variants/games_gold_seed17/tokenizer/tokenizer_config.json new file mode 100644 index 0000000000000000000000000000000000000000..9a87e3822403fb40a9524bb75edb66c6f0585750 --- /dev/null +++ b/variants/games_gold_seed17/tokenizer/tokenizer_config.json @@ -0,0 +1,30 @@ +{ + "add_prefix_space": false, + "backend": "tokenizers", + "bos_token": null, + "clean_up_tokenization_spaces": false, + "eos_token": "<|im_end|>", + "errors": "replace", + "extra_special_tokens": [ + "<|im_start|>", + "<|im_end|>", + "<|object_ref_start|>", + "<|object_ref_end|>", + "<|box_start|>", + "<|box_end|>", + "<|quad_start|>", + "<|quad_end|>", + "<|vision_start|>", + "<|vision_end|>", + "<|vision_pad|>", + "<|image_pad|>", + "<|video_pad|>" + ], + "is_local": true, + "local_files_only": true, + "model_max_length": 131072, + "pad_token": "<|endoftext|>", + "split_special_tokens": false, + "tokenizer_class": "Qwen2Tokenizer", + "unk_token": null +} diff --git a/variants/games_gold_seed17/train_log.json b/variants/games_gold_seed17/train_log.json new file mode 100644 index 0000000000000000000000000000000000000000..265131fc0a18ace4749e8ea6504d65f2a01c2c2d --- /dev/null +++ b/variants/games_gold_seed17/train_log.json @@ -0,0 +1,1463 @@ +[ + { + "step": 1, + "phase": "full", + "loss": 0.8457853049039841, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 1.6175655201077461, + "batch_question_ids_sha256": "71eb3eb01f17864e38312e7247d91f9bbc62370757ed14631dfb2992ded1840f" + }, + { + "step": 2, + "phase": "full", + "loss": 0.9267590641975403, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 2.5882841059938073, + "batch_question_ids_sha256": "04e46755a6d2d34be47ebff18d8f9b50de155c091ac7965ab70fcc86c3afba3b" + }, + { + "step": 3, + "phase": "full", + "loss": 0.5794186294078827, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 3.3830699054524302, + "batch_question_ids_sha256": "acf86d030ea4801a0d460cbf3fe3f905b2aa32111946551f8c7834c20c6638a0" + }, + { + "step": 4, + "phase": "full", + "loss": 0.7766970172524452, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 4.486405777744949, + "batch_question_ids_sha256": "5bef47fbd27d038ec388c9e7b74defd7cbcfbfe4d39a0a051e4bb5c65d8eda48" + }, + { + "step": 5, + "phase": "full", + "loss": 0.6577783301472664, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 5.422028079628944, + "batch_question_ids_sha256": "39035118b3098685d3ad65892ed079d6ccd6bad6160b7edec6be9f3e09f987c3" + }, + { + "step": 6, + "phase": "full", + "loss": 0.7534679174423218, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 6.1452732821926475, + "batch_question_ids_sha256": "3c5bbfa640330893bb222fd546d7c965dff71a2d03a77357578030927d3c6c0d" + }, + { + "step": 7, + "phase": "full", + "loss": 0.6323597207665443, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 7.139142352156341, + "batch_question_ids_sha256": "38a2c2a94d45640c15fc800a14c305f48e84ade107d3a11467a695f0d39b2838" + }, + { + "step": 8, + "phase": "full", + "loss": 0.6073539853096008, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 8.06670271512121, + "batch_question_ids_sha256": "82e99a1519afc090e56b0b35d12a5807d49925fa01d48b83e4153e8decd2d685" + }, + { + "step": 9, + "phase": "full", + "loss": 0.7351995408535004, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 8.813088703900576, + "batch_question_ids_sha256": "6955b8d432a63c90fb3ef91c2a2fdd826f24e379ec83fefbaba7c5e1a2c7f6e2" + }, + { + "step": 10, + "phase": "full", + "loss": 0.6596381217241287, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 9.590820426121354, + "batch_question_ids_sha256": "d5828c3633c79dd6ca7f3d4ad55f3b4cd720ef06583e2c10fb7c90bb1265c748" + }, + { + "step": 11, + "phase": "full", + "loss": 0.4863857105374336, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 10.467646835371852, + "batch_question_ids_sha256": "837f7e259dfd55075b1ef950e3ae3ba466eb802fa44bf0668eb1aa27786e0bd1" + }, + { + "step": 12, + "phase": "full", + "loss": 0.654490739107132, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 11.27953877300024, + "batch_question_ids_sha256": "baee215c94ec1c935eeab50a7e5587a6172042a6fb5b1a87469cfd101f18d886" + }, + { + "step": 13, + "phase": "full", + "loss": 0.5122634246945381, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 12.153916295617819, + "batch_question_ids_sha256": "d82cb144b306e38bfef9efbaa461d00bffdfaeb97e50a686af537e6b45040194" + }, + { + "step": 14, + "phase": "full", + "loss": 0.5086483210325241, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 13.073214737698436, + "batch_question_ids_sha256": "9672810ede4adc381e32df8195d738e7ad7c0ac1e7f88a35bfd7872d768f2aad" + }, + { + "step": 15, + "phase": "full", + "loss": 0.615757629275322, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 14.074581444263458, + "batch_question_ids_sha256": "0c4e89ebd825015a02ae89376a037f7481c57f84445b432b83aa29bcb6a79f5a" + }, + { + "step": 16, + "phase": "full", + "loss": 0.7543521672487259, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 15.104267627932131, + "batch_question_ids_sha256": "6e7ec78f9d346dff785e203fb5f0878d1e9d17e6b8adf219bca969d84d87d272" + }, + { + "step": 17, + "phase": "full", + "loss": 0.7431886270642281, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 16.42122661229223, + "batch_question_ids_sha256": "57f251919d9fa006aeee6ecd0eeef8c56e32dff4b02b3828d0f4bc6021acfcef" + }, + { + "step": 18, + "phase": "full", + "loss": 0.5777432397007942, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 17.445013084448874, + "batch_question_ids_sha256": "33983d483abe523f73b0852e1fbe67fff066c6be778229c8f07c2efb4837c6e5" + }, + { + "step": 19, + "phase": "full", + "loss": 0.5027384478598833, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 18.188755275681615, + "batch_question_ids_sha256": "ccaa2d32b14b5719d48d75f730b08cf1e7b38472a889edae788d72a579395907" + }, + { + "step": 20, + "phase": "full", + "loss": 0.4725775346159935, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 19.02001897059381, + "batch_question_ids_sha256": "8147d664ccf626853525d0429ca813fe6a6f278e3494e8eca72fa874da807f0c" + }, + { + "step": 21, + "phase": "full", + "loss": 0.6131691709160805, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 20.581622577272356, + "batch_question_ids_sha256": "97ddf99919459609be87bdb72407bc874134f665c32b9a52b0a8e2cd58227447" + }, + { + "step": 22, + "phase": "full", + "loss": 0.618728406727314, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 21.75770793762058, + "batch_question_ids_sha256": "1b8dfdc0069cd06a0058e663209e8c92202fd6aa1c14b195f3d9f8d353c3a4f2" + }, + { + "step": 23, + "phase": "full", + "loss": 0.3649188210256398, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 22.467512719333172, + "batch_question_ids_sha256": "dd3b96de7a4909fb4b7a2a71422687c3d322f549f346a8a57eef595ff1c07a92" + }, + { + "step": 24, + "phase": "full", + "loss": 0.5655749216675758, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 23.678164017386734, + "batch_question_ids_sha256": "f4eb619ff254a12dab492a33c2c5194d5c43414601dc1ac520baa382edec6c7a" + }, + { + "step": 25, + "phase": "full", + "loss": 1.171707734465599, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 24.559699811972678, + "batch_question_ids_sha256": "8c4d207ed85531c09999fc0c3baaefa645a6356829b698bdb9eada8a557c2dbe" + }, + { + "step": 26, + "phase": "full", + "loss": 0.8550540655851364, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 25.73836652841419, + "batch_question_ids_sha256": "23b30764010a9c65783ac79c1d94b04a8e7bf2ba89311ac8908e25f1ee03adad" + }, + { + "step": 27, + "phase": "full", + "loss": 0.6925474405288696, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 27.073857949115336, + "batch_question_ids_sha256": "84bb592e129ebd137830c33dcc203fc307bce68e9edda109a42805b80cca62f4" + }, + { + "step": 28, + "phase": "full", + "loss": 0.5988892316818237, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 27.857118740677834, + "batch_question_ids_sha256": "7930e67dab2db548df36dd7048dd9c892090a86d6345cc9055e1bb4db063c1ff" + }, + { + "step": 29, + "phase": "full", + "loss": 0.5603495985269547, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 28.998272417113185, + "batch_question_ids_sha256": "3dc4bbaef2c529d782bf99c4873f8f41620bc9d2ed3099aef7ab960f332465ca" + }, + { + "step": 30, + "phase": "full", + "loss": 0.8181445002555847, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 29.997850166633725, + "batch_question_ids_sha256": "7e45e03ff05f18483892c92841d8f6c38911d059c4c4ee1cf47b902a5c44845f" + }, + { + "step": 31, + "phase": "full", + "loss": 0.24103742837905884, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 30.897351529449224, + "batch_question_ids_sha256": "7216b45294e77a47b9be54957773d6003ce9ed66b12c1bf0e357303a13229edb" + }, + { + "step": 32, + "phase": "full", + "loss": 0.4381651543080807, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 31.888322063721716, + "batch_question_ids_sha256": "acc34f87b042d7a6213999d20ec218ccb5c83901b0d525caf643ee036f891ddc" + }, + { + "step": 33, + "phase": "full", + "loss": 0.7371217608451843, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 32.84122182801366, + "batch_question_ids_sha256": "6d63caa4f66b626fa75f5c410668c6ffa993c0c0d3794988563dc77201c6e1b4" + }, + { + "step": 34, + "phase": "full", + "loss": 0.6229357719421387, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 33.70439613144845, + "batch_question_ids_sha256": "913f29d16b823d421c059dbecfbe3807805236a92f86f2ea3553417b162a47e6" + }, + { + "step": 35, + "phase": "full", + "loss": 0.36400748044252396, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 34.50922131445259, + "batch_question_ids_sha256": "ce3a4e2ab697a15fdfb7ac26536c3c4c10b83c681f8f368675775b5725cd5656" + }, + { + "step": 36, + "phase": "full", + "loss": 0.43605904281139374, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 35.56658642459661, + "batch_question_ids_sha256": "2cc160dcc6f70acfb8c9a85d24bc521b0087992407388cc9c47c6595309a277d" + }, + { + "step": 37, + "phase": "full", + "loss": 0.7258672788739204, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 36.360741660930216, + "batch_question_ids_sha256": "88cebb96aeb5297df4cc08f102041ee18f646a64fa4a8b0ebff33330102a3667" + }, + { + "step": 38, + "phase": "full", + "loss": 0.3862719302996993, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 37.19276486989111, + "batch_question_ids_sha256": "71d355edc51db3112b5f9a173f3745a9ecaf695ec2665745f061c5162ca5c909" + }, + { + "step": 39, + "phase": "full", + "loss": 0.9101289510726929, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 37.98494727164507, + "batch_question_ids_sha256": "7ff6df7f98ede37c06d22f44decfa81e795d4faca6692b0ac62f678a7b6672ba" + }, + { + "step": 40, + "phase": "full", + "loss": 0.4805469550192356, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 38.71385194826871, + "batch_question_ids_sha256": "5e8adb0b54e22c15b5019dd7d6feaa0012be5dc6fd99b53e0df5bbecdbf917e3" + }, + { + "step": 41, + "phase": "full", + "loss": 0.935538113117218, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 39.86616095993668, + "batch_question_ids_sha256": "5aa95072cf87a762497b8d78c1abfc4db039d451ed162d354789459053617122" + }, + { + "step": 42, + "phase": "full", + "loss": 0.49528389051556587, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 40.840611261315644, + "batch_question_ids_sha256": "d2f09919774628e74fb49eb1b7a857e1b89ad1d293726283f657c59e9a644624" + }, + { + "step": 43, + "phase": "full", + "loss": 0.44372905045747757, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 41.79167284350842, + "batch_question_ids_sha256": "a070f1dbdd3203db8ba38d7e79b460e429256d3ecbf7a98dde3a8d7b43a1e9c4" + }, + { + "step": 44, + "phase": "full", + "loss": 0.6472629345953465, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 42.68901252094656, + "batch_question_ids_sha256": "09d53b9eac451c14b94ebaf12cb7046837f26e4bcd4182034d51814b1e7f4940" + }, + { + "step": 45, + "phase": "full", + "loss": 0.8967715501785278, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 43.552346566691995, + "batch_question_ids_sha256": "d30047767f4738865e545c7dc884c203db5221f676593c1d26e50765e5e0fbed" + }, + { + "step": 46, + "phase": "full", + "loss": 0.6275034844875336, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 44.54820702876896, + "batch_question_ids_sha256": "13c77a38dcc94b7e6469d21f3fee82014b392372e1e71cfea311ae90472e1312" + }, + { + "step": 47, + "phase": "full", + "loss": 0.6877187043428421, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 45.75374336820096, + "batch_question_ids_sha256": "278ca71914dd32cc3d8da8140774427d5d21cf5ba47b9ab2574d7144396f8206" + }, + { + "step": 48, + "phase": "full", + "loss": 0.46883656084537506, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 46.65573316626251, + "batch_question_ids_sha256": "74bd6f25b9ac7ec4815d468830ac7f070c9b00d4daf8374f9944958c50611de1" + }, + { + "step": 49, + "phase": "full", + "loss": 0.6111476570367813, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 47.53190662898123, + "batch_question_ids_sha256": "684a446ba692c8859234eeb6841c46eed63dde3212fcb3323001230ed39842b0" + }, + { + "step": 50, + "phase": "full", + "loss": 0.5943778306245804, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 48.927405731752515, + "batch_question_ids_sha256": "fb19c097c6373d5ce94c6a2b5cd67190546054e8a45271a06d07ca92c76a6561", + "dev": { + "questions": 504, + "objective": "gold_distribution", + "eligible_questions": 504, + "target_ce": 0.6102865659869575, + "target_kl": 0.5685604122407052, + "target_tv": 0.36740396390965874, + "by_target_kind": { + "deterministic_truth": { + "n": 423, + "ce": 0.5660289724652523, + "kl": 0.5660289724652523, + "tv": 0.3596534958512202 + }, + "optimal_action_policy": { + "n": 81, + "ce": 0.8414095543780815, + "kl": 0.5817801532902899, + "tv": 0.40787863043705896 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 504, + "ce": 0.6102865659869575, + "kl": 0.5685604122407052, + "tv": 0.36740396390965874 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + { + "step": 51, + "phase": "full", + "loss": 0.7982820421457291, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 57.38775280863047, + "batch_question_ids_sha256": "a0c683c7c2d6b37d3382b0fe3246dd2c590072f4df0ade409a30620fd51efd56" + }, + { + "step": 52, + "phase": "full", + "loss": 0.6462933644652367, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 58.42989825736731, + "batch_question_ids_sha256": "e37d3ef98903b18a591f04e0f2eb32ecaca7889d8b94bf27a915239efaf38e97" + }, + { + "step": 53, + "phase": "full", + "loss": 0.47797517478466034, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 59.35352932848036, + "batch_question_ids_sha256": "e6115e00f11a5ad3265bdf6f3fa0befc8cffdf179504184c24182f537d8738d2" + }, + { + "step": 54, + "phase": "full", + "loss": 0.5350752919912338, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 60.14188277442008, + "batch_question_ids_sha256": "a9f12f15ac261a2a8d3a75e88c6c3ceb792ef63be4831218bd3e253a8819a1c5" + }, + { + "step": 55, + "phase": "full", + "loss": 0.28265068493783474, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 61.10306813567877, + "batch_question_ids_sha256": "06bdf57889ae49866850c3bd98bcbe23ad032606252fa6cc6a7b5dbc523d1ed8" + }, + { + "step": 56, + "phase": "full", + "loss": 0.6259236857295036, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 62.27275068964809, + "batch_question_ids_sha256": "c6a4df9d7a7b1c1535e74b1c852032fa31e00dde3122a1b53f1cff0d2a047ce1" + }, + { + "step": 57, + "phase": "full", + "loss": 0.5199846252799034, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 63.474795053713024, + "batch_question_ids_sha256": "1443b68c59d6669f950b87714fb6327decb2a0f237f94a144b978e215e753cc3" + }, + { + "step": 58, + "phase": "full", + "loss": 0.5771275777369738, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 64.36713377572596, + "batch_question_ids_sha256": "14fa0328e9a5bc2f8e64d4767d023fa55d6da5d94fc181d82c21d314495395eb" + }, + { + "step": 59, + "phase": "full", + "loss": 0.47329577058553696, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 65.57084384560585, + "batch_question_ids_sha256": "388765d30297e541fb27135c767d14bb6bf1e3e25e16da27574916ac406817d6" + }, + { + "step": 60, + "phase": "full", + "loss": 0.5290749198757112, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 66.72340433951467, + "batch_question_ids_sha256": "a42e0f0f3a647dc6a305023ee938c9abfa0f06b0de1b9cef9a9afcee9a923eb1" + }, + { + "step": 61, + "phase": "full", + "loss": 0.6245139688253403, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 67.44425776787102, + "batch_question_ids_sha256": "76f97ac36a11ea376183499e1f14210d5e4816fda73a3f144573bd94368ad55a" + }, + { + "step": 62, + "phase": "full", + "loss": 0.5544084645807743, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 68.23381386417896, + "batch_question_ids_sha256": "dfa3e132366689570bd6827364d4a21365eb26c5f3a5d4835abf149c5c44c433" + }, + { + "step": 63, + "phase": "full", + "loss": 0.48251649737358093, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 69.13495445251465, + "batch_question_ids_sha256": "b17a0152430578906a129d787ae1a5a109f8209479b1463034a508383193059c" + }, + { + "step": 64, + "phase": "full", + "loss": 0.48881381563842297, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 70.0059216786176, + "batch_question_ids_sha256": "9bf273232486ef4c8fb97a6d4838855c28e107b4dcda106caf1586399100905c" + }, + { + "step": 65, + "phase": "full", + "loss": 0.6932106167078018, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 71.06514971889555, + "batch_question_ids_sha256": "9a1e4dff0ccc36c6c45dafc6115c88032070944d6ff0d5a6fde2647257076ab1" + }, + { + "step": 66, + "phase": "full", + "loss": 0.5685360692441463, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 72.04308828897774, + "batch_question_ids_sha256": "546f56d2912ed2b56823928b8ed9e61e991f2749868fbdfa3c774e5de763eb0b" + }, + { + "step": 67, + "phase": "full", + "loss": 0.6825120747089386, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 72.86048690136522, + "batch_question_ids_sha256": "440bb8716f1670fb398ba96f41644b133f6dac7405d02446b57665f6f24f47fb" + }, + { + "step": 68, + "phase": "full", + "loss": 0.43560931645333767, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 73.60489110648632, + "batch_question_ids_sha256": "7f347b564cbf507c1a37e6b443a6bdb454727112e40ad45ab49dedd7ab413c58" + }, + { + "step": 69, + "phase": "full", + "loss": 0.48012612015008926, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 74.59683271870017, + "batch_question_ids_sha256": "47cbefbeaff530de30d966116b86f461dfa61bb00cf788d277112fae42a43da0" + }, + { + "step": 70, + "phase": "full", + "loss": 0.47261856123805046, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 75.60819199867547, + "batch_question_ids_sha256": "e7c8bfc20ddca9615f073b73a810cedc1ba66a269be758c1638b9354346c630f" + }, + { + "step": 71, + "phase": "full", + "loss": 0.5614196732640266, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 76.58001516293734, + "batch_question_ids_sha256": "f7118af750855a041b00004fa68aaae33472879ff378399c4d7bbffbca70133c" + }, + { + "step": 72, + "phase": "full", + "loss": 0.4607573300600052, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 77.45826878212392, + "batch_question_ids_sha256": "3ea359cf0930d43cdec24a0d5eca7882356974b0ee7473235af99695bc232a01" + }, + { + "step": 73, + "phase": "full", + "loss": 0.6071496531367302, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 78.37541252374649, + "batch_question_ids_sha256": "c64e45ae52934e126ac7b5e6abf272e7998a24a6d2a8b5f8588a3fa2abdb9b13" + }, + { + "step": 74, + "phase": "full", + "loss": 0.5006447434425354, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 79.42598534561694, + "batch_question_ids_sha256": "c9f9cb2dae718b696a2f69a19eaf3a67f99f28b3ba10f5dae869fa60c756824d" + }, + { + "step": 75, + "phase": "full", + "loss": 0.578307218849659, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 80.1876826947555, + "batch_question_ids_sha256": "07c5580d1c8e3e9cf28d5b785d6f4a5bea33083fd382d4ad40f423d68b1c9447" + }, + { + "step": 76, + "phase": "full", + "loss": 0.476165521889925, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 81.15455393027514, + "batch_question_ids_sha256": "43d36977fea1bcfcbf04da8ab0c4d53035c0c5b6fc49c72025c018282da7e646" + }, + { + "step": 77, + "phase": "full", + "loss": 1.0061771646142006, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 82.08982643764466, + "batch_question_ids_sha256": "f2c2e30ee3fa1b95b0d71d5b2bc022d7fee379a6263a5cbc6fadb9251e1fdcda" + }, + { + "step": 78, + "phase": "full", + "loss": 0.18816663790494204, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 82.9288059649989, + "batch_question_ids_sha256": "0ed1ed699e67e3a7c515555f60b8a7fbfc4d9e0b8d712e5c330422e2cfcd84a7" + }, + { + "step": 79, + "phase": "full", + "loss": 0.6949260998517275, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 84.0813261307776, + "batch_question_ids_sha256": "4b99bb64d978903ed6c4e27a962cd9366a9e4ec3c1e985c30230c70616fa86cd" + }, + { + "step": 80, + "phase": "full", + "loss": 1.1416035145521164, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 85.16037752199918, + "batch_question_ids_sha256": "7bf2d3024d5069806bb4f345bd168219ecb2b00abfabf9fa94ba6c6025cd2a8b" + }, + { + "step": 81, + "phase": "full", + "loss": 0.5438215732574463, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 86.35963224247098, + "batch_question_ids_sha256": "466350587f1f161a8b1b6c9fdd303c8d17c19ecb3a8a44a57bbd370873419617" + }, + { + "step": 82, + "phase": "full", + "loss": 0.7272363603115082, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 87.09407550655305, + "batch_question_ids_sha256": "85e80d61fe57d0a8e1dd2aa9f608d6f7d97a644fb466b1a10201c7db0fd74aad" + }, + { + "step": 83, + "phase": "full", + "loss": 0.7460779398679733, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 88.27780481800437, + "batch_question_ids_sha256": "498f4c5921ba16ebbfdae8917b1a1e1f6b73d8bb63989d568540b3345c3f004f" + }, + { + "step": 84, + "phase": "full", + "loss": 0.5270108729600906, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 89.37272816617042, + "batch_question_ids_sha256": "6570adfc74eeef0ec2542dca9529507bf8c6d4cc164fd92fa9e410a190dd09f2" + }, + { + "step": 85, + "phase": "full", + "loss": 0.4950198158621788, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 90.22599623445421, + "batch_question_ids_sha256": "33ed1ea1b7907fffeddb1a1d91d717c1c8801752b0ab07a4ab59e3d1ea647e72" + }, + { + "step": 86, + "phase": "full", + "loss": 0.5164646208286285, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 91.35280579235405, + "batch_question_ids_sha256": "ebd9145f04f565a28ea542d6539cdec0cca276962f247fc5aec7a44966dd9a6a" + }, + { + "step": 87, + "phase": "full", + "loss": 0.46847840771079063, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 92.07326276507229, + "batch_question_ids_sha256": "0f4efb584e2406a5f129ad25d833fe1a8510477f29788a357544901fe9cb9fd7" + }, + { + "step": 88, + "phase": "full", + "loss": 0.6883835941553116, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 93.13198962714523, + "batch_question_ids_sha256": "6758984a470e7dbe6f09973a37f829ed61de5eccd6d836e239b5131856b003b6" + }, + { + "step": 89, + "phase": "full", + "loss": 0.8775817174464464, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 94.56552126444876, + "batch_question_ids_sha256": "9475cbecd78df614de186c1c0199ea43a42230556702547386af15e6ecb6f682" + }, + { + "step": 90, + "phase": "full", + "loss": 0.4405720438808203, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 95.30438126157969, + "batch_question_ids_sha256": "4483962c4c7679e7799c9db1a69a1a6663b88cd077d1cdba0228de08876f9906" + }, + { + "step": 91, + "phase": "full", + "loss": 0.70119459182024, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 96.07722216378897, + "batch_question_ids_sha256": "eef59b497d0c5228854b64433f6251f39419fb9b3df602f5096bacb412b7fcb7" + }, + { + "step": 92, + "phase": "full", + "loss": 0.5396261513233185, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 97.28353088349104, + "batch_question_ids_sha256": "b0c9356a9c3b9d88174e08162c3b4625c95b7ecc5aec32a48576aed46c916f45" + }, + { + "step": 93, + "phase": "full", + "loss": 0.5009284913539886, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 98.3207802111283, + "batch_question_ids_sha256": "1acc7a42f71f2458e0132552a29dea250e80576fada75d03bec4bc9b75f879ca" + }, + { + "step": 94, + "phase": "full", + "loss": 0.4285508543252945, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 99.06377668492496, + "batch_question_ids_sha256": "4764b0693e15f6775e078c536944774313f1b203a68057bc6a9dcd1411a6dda9" + }, + { + "step": 95, + "phase": "full", + "loss": 0.486830398440361, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 99.83737594913691, + "batch_question_ids_sha256": "c48d5599b65acd7be62bd43bb1eb151ccf6782a93c2e4700009b6b6a8487b83d" + }, + { + "step": 96, + "phase": "full", + "loss": 0.5376863554120064, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 100.77180433738977, + "batch_question_ids_sha256": "97acb22c1d88b7ee7a3677a38a87183e4f1d947004d6cc4eeedf709078d87daa" + }, + { + "step": 97, + "phase": "full", + "loss": 0.37329691648483276, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 101.53952985908836, + "batch_question_ids_sha256": "ed4a1d75bd7f4d05136cbb21aa242bc4db5ed0f8f4486ea8e9f9c32af1d68beb" + }, + { + "step": 98, + "phase": "full", + "loss": 0.34072295762598515, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 102.49611213244498, + "batch_question_ids_sha256": "fef194150154df68a06d689f833c1e4c1a1d3df55d73c2e37fcf8384a0cb45ac" + }, + { + "step": 99, + "phase": "full", + "loss": 0.5117827132344246, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 103.59738276712596, + "batch_question_ids_sha256": "f1ee5e23d392134c6e0513420f98e934b8f665ff7160f171763ff9adae2249e7" + }, + { + "step": 100, + "phase": "full", + "loss": 0.47948775440454483, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 104.33425801526755, + "batch_question_ids_sha256": "cd86804d757f592b794023cb951268f183fdfaf62c5c44fc5e62ca7d853d6fff", + "dev": { + "questions": 504, + "objective": "gold_distribution", + "eligible_questions": 504, + "target_ce": 0.6079203053015539, + "target_kl": 0.566194151555302, + "target_tv": 0.33998366396415136, + "by_target_kind": { + "deterministic_truth": { + "n": 423, + "ce": 0.5475351556056708, + "kl": 0.5475351556056708, + "tv": 0.3248628333506329 + }, + "optimal_action_policy": { + "n": 81, + "ce": 0.9232649759356132, + "kl": 0.6636355748478214, + "tv": 0.41894800161253004 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 504, + "ce": 0.6079203053015539, + "kl": 0.566194151555302, + "tv": 0.33998366396415136 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + { + "step": 101, + "phase": "full", + "loss": 0.5383809804916382, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 114.103146773763, + "batch_question_ids_sha256": "ca9a2ea8531f5c34d4a585ea5554bcd5ad60e7d2b053831b07017f74774739e8" + }, + { + "step": 102, + "phase": "full", + "loss": 0.4504055827856064, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 115.03309094533324, + "batch_question_ids_sha256": "0701dd656dc58efbf2b8c32150d30a00e6051dbd426b04288e310217b97b2e90" + }, + { + "step": 103, + "phase": "full", + "loss": 0.6407238021492958, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 115.81938582658768, + "batch_question_ids_sha256": "bb80b0fe9b7ae07234c09abf46b5a3b9c9f60f5abb2efa1d7bd1be7ac43576e0" + }, + { + "step": 104, + "phase": "full", + "loss": 0.5557990148663521, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 116.798139850609, + "batch_question_ids_sha256": "d997a8398c712577769f5faed64f77cccf80ba5aa70dac69a12a5b95a719e6c0" + }, + { + "step": 105, + "phase": "full", + "loss": 0.5066426731646061, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 118.0961615871638, + "batch_question_ids_sha256": "5e2d893d11da219372f1819882ecb9b02c6ef06dea0059699124fdd136d0ae5c" + }, + { + "step": 106, + "phase": "full", + "loss": 0.3043572371825576, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 119.01483165100217, + "batch_question_ids_sha256": "0aa28df00bdfd8780645966f39d4c6d27a6781add0a86e69b63192dae1b46f29" + }, + { + "step": 107, + "phase": "full", + "loss": 0.4562606164254248, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 120.27990161534399, + "batch_question_ids_sha256": "03514d7a4fdb12b42e965e653fc178577b1fa8c995d95bbfcb6c8f389fe1e139" + }, + { + "step": 108, + "phase": "full", + "loss": 0.5297593623399734, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 121.18794722016901, + "batch_question_ids_sha256": "ddfd43c5dd42cc7eaa3e9184e44a8ff1309eea28f08af1ccf86070312d0d4f4d" + }, + { + "step": 109, + "phase": "full", + "loss": 0.44593406468629837, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 122.25887713674456, + "batch_question_ids_sha256": "fd3eb5c3e22d6c00783a6b2ef71e422dace57af3934a8ff38da89fd134774d06" + }, + { + "step": 110, + "phase": "full", + "loss": 0.9526784047484398, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 123.05749771092087, + "batch_question_ids_sha256": "b4d295e824f7ed0925b9638915a97fe2c09012348f4b6d80efc71e3d508bc5ca" + }, + { + "step": 111, + "phase": "full", + "loss": 0.44634226709604263, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 123.83900488447398, + "batch_question_ids_sha256": "e2007f573d050afe8b43b1f9e0fe7cbb043ba234f3ad9d928db48f0f5ccc1e44" + }, + { + "step": 112, + "phase": "full", + "loss": 0.5950288288295269, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 124.63943332340568, + "batch_question_ids_sha256": "569fc8e6a96ce39ac5ac10ad63acaf581b3cfc9c456c665a119e094bbedf717f" + }, + { + "step": 113, + "phase": "full", + "loss": 0.5690838992595673, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 125.74437272921205, + "batch_question_ids_sha256": "91a88216ca7ff17530d883e7c0bfa31594091a484c27933ff5235d0834d99944" + }, + { + "step": 114, + "phase": "full", + "loss": 0.41171083599328995, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 126.86353484820575, + "batch_question_ids_sha256": "4bf067098da6a446ab470b7d5d473bd32b8a16293c0b10bfb80a3dfdf96ab7fa" + }, + { + "step": 115, + "phase": "full", + "loss": 0.6493284329771996, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 127.80306135118008, + "batch_question_ids_sha256": "7c8f1825633fb8d3f3953ec6c94d8d25d15f2c3c12d79f94d75c88f794772f93" + }, + { + "step": 116, + "phase": "full", + "loss": 0.37556980177760124, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 128.89136975817382, + "batch_question_ids_sha256": "90cb519ecf84f438a182bbdb05c9495916bcdceb91439642c475e007678c9ffe" + }, + { + "step": 117, + "phase": "full", + "loss": 0.6875787004828453, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 129.9633783120662, + "batch_question_ids_sha256": "9d63f6cd0cda7ff4fa0ac5d0801c44631167d287526ca363f3ffd918fc889092" + }, + { + "step": 118, + "phase": "full", + "loss": 0.47425008937716484, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 131.2125549390912, + "batch_question_ids_sha256": "e49d9321112e113a318ec287e7f0eb3749d664243b4b0df17225ca762269e2f3" + }, + { + "step": 119, + "phase": "full", + "loss": 0.41530514881014824, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 131.9410278527066, + "batch_question_ids_sha256": "6a0fd69c140e973d0313734692363b20bbc522a613735373f16bf7409fee0a13" + }, + { + "step": 120, + "phase": "full", + "loss": 0.39002723153680563, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 132.72213344927877, + "batch_question_ids_sha256": "584d370eed554028a8a28decad1b4ec55d6609be841077af367b27c019114350" + }, + { + "step": 121, + "phase": "full", + "loss": 0.820814236998558, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 133.9292372930795, + "batch_question_ids_sha256": "a17f6bf43045393d4e8ab4bfab06a8b4da94cb83e0e5f5524a10772b8e9d4e59" + }, + { + "step": 122, + "phase": "full", + "loss": 0.3337571695446968, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 134.6437693554908, + "batch_question_ids_sha256": "d197d2fb1c7ef748982ab7304380813d583fb7c4fb421fbf42c9eb0dba2998f1" + }, + { + "step": 123, + "phase": "full", + "loss": 0.38531759567558765, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 135.38296310510486, + "batch_question_ids_sha256": "8e077a53f43e48120af8e3a44c1fdece346a161ba63e5e94e1e7f0ee1171c1bb" + }, + { + "step": 124, + "phase": "full", + "loss": 0.6975102350115776, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 136.4440863011405, + "batch_question_ids_sha256": "35f5df3a806f895e330d9609d9fe215f56f62a7c625f7207bee867bf95ae08a9" + }, + { + "step": 125, + "phase": "full", + "loss": 0.8547459542751312, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 137.5725793391466, + "batch_question_ids_sha256": "873e668aee9e67a73ea7baf22490f05a776bf5d04b1166d5b2c2ea7dc250ccbb" + }, + { + "step": 126, + "phase": "full", + "loss": 0.6699497699737549, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 138.79905671440065, + "batch_question_ids_sha256": "68f8c0f3383f82951a82143972316311a5ff8c250647d23db72c5b2ba1ff1f87" + }, + { + "step": 127, + "phase": "full", + "loss": 0.743767574429512, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 139.96663841139525, + "batch_question_ids_sha256": "e85ab1fa9eea973dd7f76f9db986114c83af15353e93f85a24ab11741e30fb49" + }, + { + "step": 128, + "phase": "full", + "loss": 0.3533981665968895, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 141.286115068011, + "batch_question_ids_sha256": "b536c26d25f2e517c4f37ca84501839d9ce5e34e1935fbb6cda4c6c1f7fede19" + }, + { + "step": 129, + "phase": "full", + "loss": 0.6029391214251518, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 142.46502018067986, + "batch_question_ids_sha256": "7001b48eaad499a7aa03c79fa01bbcb9c7b71f038e5294deae1069b6bfe1c2d6" + }, + { + "step": 130, + "phase": "full", + "loss": 0.3786966558545828, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 143.28406373690814, + "batch_question_ids_sha256": "d3a61a405c35262acdccb44d96e7b94d53decab87d9092f83fcb44ec3206a773" + }, + { + "step": 131, + "phase": "full", + "loss": 0.5346395932137966, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 144.30479943566024, + "batch_question_ids_sha256": "54a9c0ca14ccaddee6aa898c5f511046e3826fd238a54f36575eb26881394a3b" + }, + { + "step": 132, + "phase": "full", + "loss": 0.45413441210985184, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 145.4113561231643, + "batch_question_ids_sha256": "312c310b1fb5ccbf8115b9228a6af966c32d20363a8d05c1bc65a9d31893e07f" + }, + { + "step": 133, + "phase": "full", + "loss": 0.5446238778531551, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 146.4893439952284, + "batch_question_ids_sha256": "e496941c457d59498818e83203e8072d880bee77ecdaadae847a616084b064ba" + }, + { + "step": 134, + "phase": "full", + "loss": 0.5164259895682335, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 147.28010689094663, + "batch_question_ids_sha256": "901b6a776006e4b80ca92b6260b2b8cb910ea19b1d535422da631eb894abe301" + }, + { + "step": 135, + "phase": "full", + "loss": 0.46057215332984924, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 148.0218255566433, + "batch_question_ids_sha256": "387851c5f1678867d22336308b62691d5a88a8716c34cae9defdea88f91bcb91" + }, + { + "step": 136, + "phase": "full", + "loss": 0.3152983635663986, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 148.7750539202243, + "batch_question_ids_sha256": "b0d587630845dd4ad4182266a0026554f493ec673ed9d35d15a256113a052ae7" + }, + { + "step": 137, + "phase": "full", + "loss": 0.6339995414018631, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 149.72458379808813, + "batch_question_ids_sha256": "2c4031b28f2741376d6b9db730150bef1836861fc24869609e30e2650ecb4ff1" + }, + { + "step": 138, + "phase": "full", + "loss": 0.38320328295230865, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 150.71331149805337, + "batch_question_ids_sha256": "5aa1c649a775ad73b24a3317de5396bc81ea188dc4d9c39fc84f48760e0b73f4" + }, + { + "step": 139, + "phase": "full", + "loss": 0.6110421605408192, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 151.48428997956216, + "batch_question_ids_sha256": "027e2fb10a074ca00739112299fd18a58aa56b68b11158d55d83074df9e66dfc" + }, + { + "step": 140, + "phase": "full", + "loss": 0.817102700471878, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 152.58210559003055, + "batch_question_ids_sha256": "5794b237a04d773d776560f43e44887520b46939d7bba10b09f5ffea11dfdbca" + }, + { + "step": 141, + "phase": "full", + "loss": 0.25834623631089926, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 153.55816171038896, + "batch_question_ids_sha256": "23d0fb5796158186fe209acec4f35212bb44956ad5d04f74b53e078b7e1ded0d" + }, + { + "step": 142, + "phase": "full", + "loss": 0.32701360061764717, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 154.3068744475022, + "batch_question_ids_sha256": "57115832176546b32e2f6548614021d5310c2c1c247c87de183702ab5eec416b" + }, + { + "step": 143, + "phase": "full", + "loss": 0.31050091283395886, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 155.2095251576975, + "batch_question_ids_sha256": "7480bb4b00af69a66fb177005c98a5da1cc843dbf68958f519868fc7857fc564" + }, + { + "step": 144, + "phase": "full", + "loss": 0.3840100672096014, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 156.1972500877455, + "batch_question_ids_sha256": "e987ca1cb740f2842f80ee77c38dfe958f4b193bedb0fa171cf6062cd6de9aeb" + }, + { + "step": 145, + "phase": "full", + "loss": 0.34445524774491787, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 157.06585253030062, + "batch_question_ids_sha256": "3a603892e4036b156fd11c02eeb989d9b4d6f6bbc0f3a7f0443fa0909fbc983f" + }, + { + "step": 146, + "phase": "full", + "loss": 0.6161299049854279, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 157.97192534897476, + "batch_question_ids_sha256": "111e24c400e623bc4fa0ee2e3d056aedbbb34da5a77caedb5c7dd24d2edc13e1" + }, + { + "step": 147, + "phase": "full", + "loss": 0.5337724909186363, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 159.01290085166693, + "batch_question_ids_sha256": "8bdc0029d631aefcd201c7376baeb6ea94e48aefb8a16bee68d60d1026f6fcc4" + }, + { + "step": 148, + "phase": "full", + "loss": 0.40094537287950516, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 159.77310836873949, + "batch_question_ids_sha256": "e426158ccd076062adf85969cd4a75b216716cf5735f243352d4c7f6a1508863" + }, + { + "step": 149, + "phase": "full", + "loss": 0.5570383816957474, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 160.59478989243507, + "batch_question_ids_sha256": "ee638f0f7b35e541f440118e59137fed22bc01e2fba353c3aaeff6dbd740b569" + }, + { + "step": 150, + "phase": "full", + "loss": 0.47629961371421814, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 161.63340393826365, + "batch_question_ids_sha256": "5bee948b48b7bb5889dfb84662429a3bcb352f9fc1453658bfc3eab909f26474", + "dev": { + "questions": 504, + "objective": "gold_distribution", + "eligible_questions": 504, + "target_ce": 0.6113742163886272, + "target_kl": 0.569648062642375, + "target_tv": 0.33253138166134766, + "by_target_kind": { + "deterministic_truth": { + "n": 423, + "ce": 0.5666581907745619, + "kl": 0.5666581907745619, + "tv": 0.32191759365659844 + }, + "optimal_action_policy": { + "n": 81, + "ce": 0.8448912390398573, + "kl": 0.5852618379520654, + "tv": 0.38795894124170455 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 504, + "ce": 0.6113742163886272, + "kl": 0.569648062642375, + "tv": 0.33253138166134766 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + } +] diff --git a/variants/local_atomic_seed17/backbone_config/config.json b/variants/local_atomic_seed17/backbone_config/config.json new file mode 100644 index 0000000000000000000000000000000000000000..b815be77563f3eef88baf4b4c8ae59706e618ad2 --- /dev/null +++ b/variants/local_atomic_seed17/backbone_config/config.json @@ -0,0 +1,63 @@ +{ + "architectures": [ + "Qwen3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "bos_token_id": 151643, + "dtype": "float32", + "eos_token_id": 151645, + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 1024, + "initializer_range": 0.02, + "intermediate_size": 3072, + "layer_types": [ + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention", + "full_attention" + ], + "max_position_embeddings": 40960, + "max_window_layers": 28, + "model_type": "qwen3", + "num_attention_heads": 16, + "num_hidden_layers": 28, + "num_key_value_heads": 8, + "pad_token_id": null, + "rms_norm_eps": 1e-06, + "rope_parameters": { + "rope_theta": 1000000, + "rope_type": "default" + }, + "sliding_window": null, + "tie_word_embeddings": true, + "transformers_version": "5.17.0", + "use_cache": false, + "use_sliding_window": false, + "vocab_size": 151936 +} diff --git a/variants/local_atomic_seed17/best.safetensors b/variants/local_atomic_seed17/best.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..d41ade7dc91e27c6f64062c331dbfe48db5aeaf3 --- /dev/null +++ b/variants/local_atomic_seed17/best.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b06e4423861f47da7ca53ad9c23a26aebdbfe79de3a238b36c2954ae9c6abf0 +size 2385039280 diff --git a/variants/local_atomic_seed17/config.json b/variants/local_atomic_seed17/config.json new file mode 100644 index 0000000000000000000000000000000000000000..9f4149f52af007d0caf230c528163aba3f905240 --- /dev/null +++ b/variants/local_atomic_seed17/config.json @@ -0,0 +1,69 @@ +{ + "input": "data/local_maze_v1", + "output_dir": "runs/local_atomic_seed17", + "model": "Qwen/Qwen3-0.6B", + "revision": "main", + "init_checkpoint": "/home/rwang/openjev_codex_20260917/runs/v3_teacher_coords_multi_seed17", + "objective": "gold_distribution", + "loss": "ce", + "reward_samples": 32, + "gradient_checkpointing": true, + "set_head": "attention", + "steps": 300, + "head_steps": 0, + "batch_questions": 16, + "microbatch_questions": 4, + "max_microbatch_tokens": 16384, + "eval_every": 50, + "max_length": 2048, + "seed": 17, + "backbone_lr": 2e-05, + "head_lr": 0.0002, + "head_warmup_lr": 0.001, + "precision": "bf16", + "disable_native_triton": true, + "validate_only": false, + "self_check": false, + "schema_version": "openjev-decision-pipeline-v1", + "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca", + "initialization": "local DecisionModel warm start, fresh optimizer", + "data_sha256": { + "data/local_maze_v1/train.jsonl": "502f2840f3e15ffac76554f35ef6c2e41c6e25f58f77b5ee3a2f1196cba0c41a", + "data/local_maze_v1/dev.jsonl": "bd7f9d4402112770f571210677e329e8673a38caa6bb798b744ee47bc793b9a5", + "data/local_maze_v1/calibration.jsonl": "94a233e71eaf3460c939bb0b4b147a0a951931001097070ea7ac8b4b78f39ace", + "data/local_maze_v1/test.jsonl": "be7783c395715f9fa2aae7c793c4e5b7c4b1dfe3fcb0880544b577e85b8f3b10", + "data/local_maze_v1/ood.jsonl": "e6772998a189880c780b6408c0e348c77324a28204e390767dec54f475d61f4e" + }, + "deps": { + "torch": "2.14.0", + "transformers": "5.17.0", + "safetensors": "0.8.0" + }, + "gpu": "NVIDIA A100-SXM4-80GB", + "parameter_storage": "float32", + "forward_autocast": "bfloat16", + "parameter_count": 596250498, + "train_questions": 576, + "all_train_questions": 576, + "schema_counts": { + "records": 300, + "questions_by_split": { + "train": 576, + "dev": 192, + "calibration": 192, + "test": 176, + "ood": 64 + }, + "eligible_by_split_objective": { + "calibration/gold_distribution": 192, + "dev/gold_distribution": 192, + "ood/gold_distribution": 64, + "test/gold_distribution": 176, + "train/gold_distribution": 576 + } + }, + "loss_description": "per-question proper loss; complete candidate normalization", + "rlcd_candidate": null, + "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection", + "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing" +} diff --git a/variants/local_atomic_seed17/summary.json b/variants/local_atomic_seed17/summary.json new file mode 100644 index 0000000000000000000000000000000000000000..5e1ec1743974e5a416b24462a580fd8e8b726a5c --- /dev/null +++ b/variants/local_atomic_seed17/summary.json @@ -0,0 +1,160 @@ +{ + "best_step": 300, + "best_dev_target_ce": 0.48294000757494554, + "selected_on": "dev target CE", + "objective": "gold_distribution", + "metrics_by_split": { + "dev": { + "questions": 192, + "objective": "gold_distribution", + "eligible_questions": 192, + "target_ce": 0.48294000757494554, + "target_kl": 0.48294000757494554, + "target_tv": 0.31825824177139705, + "by_target_kind": { + "deterministic_truth": { + "n": 192, + "ce": 0.48294000757494554, + "kl": 0.48294000757494554, + "tv": 0.31825824177139705 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 192, + "ce": 0.48294000757494554, + "kl": 0.48294000757494554, + "tv": 0.31825824177139705 + }, + "observed_outcome": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "calibration": { + "questions": 192, + "objective": "gold_distribution", + "eligible_questions": 192, + "target_ce": 0.38027106708044384, + "target_kl": 0.38027106708044384, + "target_tv": 0.26135403294069587, + "by_target_kind": { + "deterministic_truth": { + "n": 192, + "ce": 0.38027106708044384, + "kl": 0.38027106708044384, + "tv": 0.26135403294069587 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 192, + "ce": 0.38027106708044384, + "kl": 0.38027106708044384, + "tv": 0.26135403294069587 + }, + "observed_outcome": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "test": { + "questions": 176, + "objective": "gold_distribution", + "eligible_questions": 176, + "target_ce": 0.4404518677886115, + "target_kl": 0.4404518677886115, + "target_tv": 0.2833239381592964, + "by_target_kind": { + "deterministic_truth": { + "n": 176, + "ce": 0.4404518677886115, + "kl": 0.4404518677886115, + "tv": 0.2833239381592964 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 176, + "ce": 0.4404518677886115, + "kl": 0.4404518677886115, + "tv": 0.2833239381592964 + }, + "observed_outcome": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + }, + "ood": { + "questions": 64, + "objective": "gold_distribution", + "eligible_questions": 64, + "target_ce": 0.44655846661295645, + "target_kl": 0.44655846661295645, + "target_tv": 0.31297003501627385, + "by_target_kind": { + "deterministic_truth": { + "n": 64, + "ce": 0.44655846661295645, + "kl": 0.44655846661295645, + "tv": 0.31297003501627385 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 64, + "ce": 0.44655846661295645, + "kl": 0.44655846661295645, + "tv": 0.31297003501627385 + }, + "observed_outcome": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + "temperature": 1.0, + "temperature_fitted": false, + "training_seconds": 242.69976669549942, + "max_gpu_allocated_gb": 11.958739456 +} diff --git a/variants/local_atomic_seed17/target_audit.json b/variants/local_atomic_seed17/target_audit.json new file mode 100644 index 0000000000000000000000000000000000000000..affa0daa883773ccc59305a26c79afdff9967277 --- /dev/null +++ b/variants/local_atomic_seed17/target_audit.json @@ -0,0 +1,20402 @@ +[ + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:997f24df6e0ff0d792bae70f6ce262daf217df557b71ba6c045c8ce59da0ec63:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a156f72ec902c0bf149f9f3f46bc4a37f2983b6d3cbdef7870cb194d1fb28144:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7914c943d451ab296a3a0fb1206d5a549999ac823baf941ad1a56f165fa7b766:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f76a0900dee53128a0a1622c640c2a38adce881eeefd35f71a44092b1bf1b83:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0f41dce251a972ad27c9a14d136aecd80915c64103c0f552d1a1f469324aef0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c81c979d8e8cc60ddde7c6dec3bdfd2129a62fa6720939f763e21a033b02d4a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0bf85781b941ca7c9257a67c8431dce83d8bc3933b6b169fdb01a85a612add13:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ae81e2e92f7fe71fc8586bdea1d9c5bdb05a792af58179725eacc655bbeb1703:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a62807539046c32eebd3fee650aaeb80e0f9e75dafb4965573228dac2cc5011e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:deee330a7afa16b4977d6169037f3e22272cd6f9516f760841408d1c7bb6def3:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc6eb25fd2961ae09eddd238b39ac4f4c650a769d2c9621b3b0da8b38e18d903:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf168ed977ee7d45332928459509f7ed12589125affca64c2964e33eaf638a81:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a5d98b11436fcbf3466281148ba434803ab3e338a4f6e449221f833144ba82b4:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6073358a80f3a4710e2bc268bb42b4fb57d925b8230c6034b27c9294d93b0b5e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1a73860a9d155a3a8fe092a8f60550cef5e675576796ead21eee6207b022d8fb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:69e43d8c80957f04d313cfa0704f48a8b8f9ed66c767772b16e9b8c4a219c2b7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ec42698d9d7d0380bb85477f2d6a3e8b352f47c0972e3d9f7836cd1a29a5e23:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:60331f648cfcf0437235cab789bf180d67109ed7da65c8cec2b2a7e674ea171a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b392b0d5e5acea8ee39ca911e4a4a1a3815cdf5fe118e8fba68ed5aac9d68fcb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a61137fb698f7ad43dfc35a49211d5956addc1864cc16ec7ae3362539cd2fb7b:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a4b2bd71b8bfdb8ac532bdbff9ac917c9a71f6f4f8baccb5c6fb36ee27a162e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 171, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e1b4ffaa4fdae0d128f9bb65f6576cf8aaa3ae8aebe843a41fa21928c380ea54:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e89593a6afe69f91d7bcfa9b58e54f26652f75af09a81e7858587ba6234565f9:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a3b55dc7fc43f4d711b754b2e02da0aef7655e1c84dae7b9f6d69d3db497adbe:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7e399cc4fca77a78dd8af19be5c4800e37da3700d449c7904e758544da132ca:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8fcdeac7fe6bd393bd99768afffc475dbd46ea671cbc355ea1b6f15421833be:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc3c59daa77b24c97023c268b9002bcbb643e776533eee196b31443c2e030b09:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:928b0211a6291c4e2c3c5fd46ecf9af014a66fab48d1de8e6960457378834e4e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a8396e7ce966cecd8f4798aa722c08a9fcd19dc4f9e31b9f5e2d6eb6c7ad0e56:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:201438510263410444b4fdd919d70a6594fcad37f390a13559b07a3d11d53511:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2962a3a7daa58c1d2887c40df2ee33a3f1a3d67824a29730669b8603a01ef305:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9ada2d33d6454359cc4137bc8e206e82377eead7f333ac904149880e47404854:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9131370985cf07172713032a3407214750c11b15dab5b3f89490955f582724a8:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:43fde149f352904a6ab39b9f4d4002c65fa87fcbeb303417825ef577261324ff:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:35acfb00913a0ee5529f1a737576bcd66f60c2606bd354eed21cd3ad2d487f94:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:82dbe0bfdc1ce108b18f9d2c7980d1cb133b02c975b301d650a764ef57aef7d0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0ba229e0540e2ec7d4d2cec8f972ef4e9f55caf27acb14d68bbe566fdef56b37:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12a4ef4c1e37ef85d4b352a95ef2eb28b07bce3baca5374f8e464905480850d2:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ded3d07a7b65491265913069800db4f8992953aeb0e874c2d456c336ab75aa5:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d9e802a242a522be386189b7efec1f1bc2889247cecad2c6b15d0c878d322c79:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dd0e1861cded290e725659957a2bc2d02ef53de613c332f1582a852531149359:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:04d1804034430f8ba9738b15948c417203e12d755563ca4c60638ef895a8cc38:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6e29f99c33825818c536f664f088625358a71508a9bfcb3a33b51ef95e049f44:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:889bb1519d4dabf96e3e32d625e1076b716be2ecf5dd1c70ef253a7abca925f0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dc33dfa057079734423e123890619d5cd55a37114a7b9390b39c384fde85cf94:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cd23ba13b086cd5ec018204702fa44bf68adb2745e9b5c4f5d6eb2d65fc37df:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64bf4bc617d8f3abc81176367f439cd3e827a7469f165c80e0505221cd658586:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 184, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 184, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 184, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63f506bbaca4cb4705829a376c173352a88df3daf92cb16abc2b7a52399d23a4:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 184, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee9056d425e26b5da6079ef922c7a8e45e88a5e88fb2b749f549210892fca629:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:53218c1a70a5acc03138432545118e2b23591d13c3b8117bc18f219a9c6bf91e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00fde7041f5fa336a8ab83e94d166b093ef67edf3779e0e714d03fea563391f7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd75dcba7aef4a31ff5e54146458d0859a0c23aac148a133d9e88340b0f93b87:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bfad818b3f69c3286657ea934a94424170597268fa5fddc5e584d4e35c848b62:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0295a078a58faa63aae7205aa6115629fcaad24ef7a22de115c1a43a1920b923:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:060dbef4fffbf137514f490099181f469ded48438a2d681015a117215f9d33c5:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:54b4d3a8ac837162577cad5899930fdc3e10b118e23fb4af7d5ea2239b57f444:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4d68de96bc11e6c8cabccbd4150a7523283435fd9a3a36c91a6b57425d913dcf:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e59cfeea15ac1da91b4434f023979c8af5b5dbe1cd219d0bbea809b33fd111d:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c853b21a30428fbde72a533bbcc946584a2f6f20c0cb20e7687ea249651b4010:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f651938e1f6e26304940cbdd5211c2b645b96b9f67037b7e8734a8961d0a9004:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47ed3c2abc1c477dbaaeaa56561deb72acf4f7fc2121a7c6c5797c437d79bfdf:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9436ed512f2646ba7e447d72efc45649488d83d09d6608f91b47dadc0b979c99:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c184f4de1ce384d0ead8678240538b19a82d08eb8c46d6ae0bec543fbefd70a9:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c7f636ffa6800d14523e33d0c97c665a139efeae92e342c1c5ffdbffbb890b82:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:72562853e0f48e2e1d3728452f4fd64b5f8981f2ed34e799092c052775cdc8c5:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:db5926f0759794832e41ec1c04c8ebf611d6eeb6914dc589da9ef43ef943a39e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d89149fd85d1fe48f47a8a529307b84915d2778ec278ad40d10fc28aa6f8d3eb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0e3e0c360780a0dac85c8fde0734e1df06035c50b35a3b3450bd97e86ada8f95:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:17f9eb320f2c59f815f114f4109646008589d83f15f4f118b14eb4e008c7af19:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ede4554fa7a656d3d069e403605906f6406b11a7d7184a94df4e27429a0275b6:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:21bd093ba5b74efff880ed25745c0f94956c21ad5f2dfe3f34b6dcecac1b160f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:af1fa6b8809506545b49384dedce0e8f952bea66bd58fb32d50cc88d8619a02e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d5d91f9afb9aabcf0668411952870bb70c7adde491867284e504f2548e1b1cb3:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:76fe4041417fb02bcb957761c7df9b9675cff257707ad971ee7c466cc86787e1:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0faf23c1f21c798e644a17ce7efb7d509b29cf040c3f2bd991fe6c15f50f391:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67bf3badf319de5c5293ce1dfe667ff9723d4debbf7b36f51c16e2c9e35c9ed6:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3b0d9a8f6636940618be8b7d87da3ef57bc4cab66522cc7def0f16102d712122:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:179e50b6d2d021b16bff4f005bd0de0ca3d6627f35fcbccd894c2c06ceeff8fe:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b28c420dd7bd5424f909ef366dfa20b2fbebcd62dd673b08d47be0d90f11d92d:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91423bf6d82f2610f90c043877f4c5af077d7270d08370caded167ffc56c94c1:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:03e869a5831437c6cf1698cae026593512bbef7e308e08d26481ed52487d749f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ebb9eaad9fc7f1be3f8362dca07d0f082185c7c8883cace33ff3e65fa80c7526:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cf24f2cfbbf24f72edd3e073795107bbf350edaeeecf03d179fff860191890f0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f5ab78d5b297cc6feb36964919880e518c5e0a0b09f6501afef2a517cdbea0e7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:681090578c31b72840d3969b6ed5d2fa9837b7ed6b200a01e4e6054e58f3af31:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4ee210a3a31dcbdb9dd11761423fd7e1d20928572f2350ac6f05683c19cbd530:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:da5a9a8b850c513d61cc25e93f4ce207011b8b06e791961373ed8905a521b45d:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4c86f1f531845efb71ab0f1109ad0490b8d00fae2c7165c3552038eeb4fec246:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:98fa6fa046848d67fa32e3bb8f9bba13d294e991afae0d6267139b086d9c1b8b:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c563fa90ccabb41fb7107cdb5e94961e7360483a59501c95207518c9938022c0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d273b46e3130f12067bf93e7138592edc1be3b0119a1b1ba001e352a4a26fa17:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b7225f889a00f2dcbf98ff302eb1cd530d7574ce9e05c6a91223bd291a8afb8f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:be514abc2cf4c709b6798ae21418e3d38de77ac728cb92ddf9d0798f70e6f902:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6f01f63c2479b9449b1fe05a824d39054f3d8b4ff4574eb01fde0c9f8d1c6278:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dea8e28e36a0b390368e4da291b2f388a07c89e8f942b6a8eb8480742ace43a9:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:28afae95a6b3b628749e859a2a4640a23af34c0d11c9fcc3ff0b746b57ab3788:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0a4d156c6568e99322563c8f26f7bcff86ea918f03dfaf81a51bac8bbee5f2de:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5c54e02dfa19b5ced443d62cc31267d0527498f5f57fe0900b78291fa6da680c:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e9bd8438ae76018b857808515fd08f55f20d9274b6bf3dffc1aa7f52ef0e3adb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2e11922ed1b5408dfe84f60f10175ed3498e15c6af729d9e8929bd7909bf44fb:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a17a4ac6c6a215d41b6995772cbb91f7963db3765439b34a7539573ad30b70aa:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f60b72150300f75768101cdc93c8840c3af9fe5c3dff6d846bd67f78f1896d3d:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df6aa5176de7c9ce7ff3f4a51d0088cc0a39d7f5e57789f591f04c09fb83614a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:432b1879e86ae4f33e5a427d30e2bcf3217d8094a75b4d6829a7c40a34ccf008:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8c1554a94bc88d572348e487fc76116a949e64218fcdf2a50bed7fe66089cf0c:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e77404807cb13b2c741524553932259f91f09686ebe2b979fd06a0eb3e62ad72:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:05b1ea15d4d71585c1f8def16533bb527eb492725167f346df8e8846fb5d6a90:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23a5a45fa191fb5f43ba343361103e522225ce553c43c98a920f95ad50fa65c9:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9766067f38678f94c6a70c864d1ce73962a02131e4a97618634b1b7e6ab810a3:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b3225879428b5a4295529c9cc6612a2f07119ef404f794aa41461afa0144cb0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edd886c86f13fea5e900397752221a8db171370bfe3f16a4c6eaecf41f6215d0:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c37203a0665cb2d8e4593eb2b922a44bce11adb18ee2f53315c583e016b3c128:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8cac73c0803686a63591ad9e12e17b8bedf40c5e372a53192fb726f78cb542f4:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ec30729a6cb8f33e35bf70636212ed244dd27c5592967c5b774509d545ebd2e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9fe69fa9a58cffbf07be4dc73089ff0c9428935f314d742214d517b225b41dfc:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a573680e490552d69614f3b04c732f990416ec7c3b020059499994b9f4d434a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a888d6db4e1eaebb41e661bfbd7dbf05d122d345d23967130574764b7a2e395f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2c4bbc1d884e38a42dc7db7500bbe7c76b461c4239cb2fb0e9d05f61b02c723b:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fef75eb7c725d9a9a71d8b2619a1b72e6b9b2fbcc286c2243a558e8fc03f07e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:339ba3345f28b3eec6244fa9d80ba769f83d9678ad82ec1443a3c93e90abba8f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:64229d74ab934fcc8bd9f40f89ec2c04babf4cee73baae18317c27d9840aadfa:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f050ac9ea06fd5fd487b0c56b7889013e1bf4f05d6491bad219f0ed4d4bd882a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:55345ff14c8f638a4167bf024563ed73eca278a14077310c6d0f67bc0a9938aa:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:399055e56c7768655d377406cfe1d2c12a8a17019206092618fb651a391b5c23:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3a7133775818b0b94ba4166cb16915a6dce124a1220596dad15dba7f0e4d5f49:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7a86f955f1f227bd28af8b3e5e33138c040f2315684c66609429873475b5ee1e:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e2f117dd1c88d376c3a0fefa14d5e492247bf86133fb28f5353fba642c10042a:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f1d8ff515b54efa3507362191537c102021c6e96776cce45e3a20b8896e0acc7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0b4c1bfcbdf14ec92a9ef31ff91d780ff4eb8772bc291a29b48e954e3e860d4f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:22bb8ec9f8d38b63b0f203987658844219afb2e1a2bf49a7207a3b19d95652f7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fa5aa024241908230e9ac8dd8b4f9a2724ba39757a44d6c1ed69fa2f26375f10:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f097412d8a3720d8b62f541a241f213f2b1cac277f0ef64e1c36e0db3442f5f4:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5971f23fa39b0edee3a9a5681d5895f61e8c49aa50309784764e8b82f44973b3:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:dee049a7ddf4df6072ca92ddaba932eb0f4050e96dc5f29efc854d9e36fe10fc:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:81df64af43c06518d46d989b905c36f47dced4895a3025e97f3f603dc7e5b156:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c96a7fe898cd1f83ba10ceb836911134e47b1d4b3d0d837b000caee25ad1d7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:eeaf00a83284c1e8ba21986c4386477c9456518bfd63aa9cd439edcc40c684e7:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c0e326e288bc0952603903d477d15fd73a8cea4c38dd5e250fe9e2f276b3914f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a2a5b253df38564764731053dbc20952db101439fe14137ce958f2d315ecf7bc:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2f1e535c95b75855ed639833fae5d2067d9075eebbb9340f37275b0569ff67e2:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7ca773c3b466bba37cfda3ceee1df5c892987c7d2087c9904f14dabc7d86edef:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2ff8856bde15e389e6724fadbb54223772f06bec226ecc8bb9843d79706e889f:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b91abd99997e289a4e8392003e2771589615ba7e23c3b0bd953dbe3b2b4f7637:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:clear_east", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:clear_north", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:clear_south", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:569d718eb5dc278a72ddec112b86f271f9e629f781c0bb8b1f7d499d1c9a453b:clear_west", + "split": "train", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6fb4dcc50486636d4948fde2f4d412683e3cfa00358fd192bd7c034e1f463d26:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66295632a07890a014ef1a0a7d09c4263873e4bb54bd9f47642ff31a6608e407:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c748c0659ab1daddff3b4a4bdf1429bd2b67bcb5c83518b0cf9d2c801e40c05:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3cff3f6a2e01c0db52092e4cafaf0d64e0113bbffd5b373ae92c987a9ca45ba5:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a4c2210708f8c58156f7f610bea3be2a2731b45733178010621030a1615d4226:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7217ccc385d004f3fa640d87b6ee910cd47ba0a33a298322b9e9d274166e3b99:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adde3b834e1efe6daf72eb2558fb6355692cedcd35fb83d23b76095f91077bfe:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b391f1c565c5ccabc055d971eb4af6377da33c14a0e83657d651976417daec3f:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3f5a6cdb0675b93bf07c6964cc18386b4da64c3a5c4aef6520e0e1c7a5062851:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8396eb53c9d848e6def1a950b79a05f1b8ddaee2ee4f09dae7d055ce061274bd:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa844939d58fcef3c45b3b351a9d11bec6f8bea09a0ddb96c6bf21253926b79f:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aadede48f406712c45f9fb8120cbcf4b15688854e0cd67a9be2c49c744ffc8b:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f741f310ec6561f4748181288d33327f95244be654ea72d58cd867a4edc69ad8:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0c0f55a66ea5d2a2667050f7ea4b0428ec40d0e66f079f96b0864a1a76d101ea:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0aefe25f53fb9583893d46a1d196e6c6373b7935177652f84735336dbbae1f50:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f2d098844ee6f845ab64a9af588f8b3e3aaff993e07afa3bb55615ad406b73e5:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8d8016b5862d64311aca86d6177e635cfc29775fd7814225cc66b211ef08b2f5:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:62779fce1feeeb35740b33b81f770618c607034b98e2c84183710d344dc4d9fd:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8906117a807887d2462092d1298595228172b42cb628ff6978e113e5c28487f:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0fd3fe1cdfe0fd7d6e6f0c7ed5b1e669619f817fa2b7e4de04155bd7949877cf:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:86c9683eaa0d0f6c7282d0f748da6418bc0219331d87e2bbbe9ab4639188194e:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:00eaeea3035abda2803f177d45bfdc6cdf01887fd43e7d85d33e47652a4b74e8:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:adf9b524bd83702ad713cfc92ce34f079e91a4f8265e401dfaa9e4c0661884f0:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:662ad6f2d058a3af38c9b2ea1693c95cb0989697fcd5640fe86b9794dbebb089:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d920ceb59bc0688d952acb6728dfb10f1a4f68d814ab5f2957f0414e2d92546b:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d95a893b606078c6641cf966544acf3027e6e5094a5029ff2e952d923369a292:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fafb24d7b73d48dcbeda38e9f04734e009c713e0d3fe983d3c62aea37f5a5a37:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6dfa14e7eb68f84758c6f589daf8e502a6b94485038323dc16b31854f57f47f5:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b4ffb414f0a54f429c7e3056e4cf1bc893073d8a92280b1bebdabe22c10c44ee:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c31f5c0bf151c21995c814a6dc567a878f212f87cc505e520336968a1d630ad7:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f089f1cac8af2b9a23c318089acf6a014ef54e86734ab582ebf1b91fe46fe367:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6423fd7020c9462a7350100e45ac387938da5919c724a9a704aa668cb0f9e7a:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75b4f52c4470b521e552638cef73e40e411fc894ece33771664b220741ee30d8:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3ddc36c87c85606e452101eefd24ab3277cb642ec96c5ce5fc36803ddd2302b3:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f3216bb24150b09d4f5384f20b2c4e8777be1fc4e8aa38b73702eed9b539879f:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:922ce09b6535d1767842b7e9607fd9ff542f0933ba916db5b3b780295fe07e39:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c3a492be74ed1355fcb6f4706d054d7e89aca4cb29247dea47ec1021446084e2:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:93be99ab70deff23de8832c5a5e4158c993094204760949873a2698ccc5b54cc:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fc35519c4501283a9180c6504885a9ca147cdd7c0f6a1046349ce5daa84bf6dd:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7b663c3e688a8cb18c6af689b548be79bbf6a03e90e33a8841733a7dd3434c8c:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bb80c9548f935619df9158b346c6bd9c03e35fb8dbf5c050c6068acd04386a6c:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee0c173ce5e03d5bab10b659c997de81ece20e0f9aa4271d719484325abf8d51:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c127eb5517ccfb687d2b4e40cd12c456dbcd29463e77e689c3592254ec686559:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5695eb47f3e6b048e4ccec80c48d2d4bd6f85514a75850cd65821c242a043213:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5f2f84fffea8f6d3f0e6abbb17f71c1548d5a4f5bccc65857fc4614f350db7a6:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1872cfabd3694bfedd7eb043466a8fa030c7ebe5ffe3ed0d544c8db99eb0fd94:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8ff926e204772eff3bfece3337b4cf6ef2f287523f2f4a3458a51ccad7baeffb:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:clear_east", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:clear_north", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:clear_south", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1f154912f4645fc29472b920805b410468369b35b98674ddfd3f9a5c75df51b2:clear_west", + "split": "dev", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b25341cce6057cceda31c29ca3f4573fac85f6a7766a6011058af970e00c2ef9:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b544880e343f7d179dc3bc072c9cc7b7a4a13be9ee8cb3f7249c5037edc045c3:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bc9e33bcbd30f6ca02b4fdec266200d88b2318a9baf3a950536f6644f31c2e41:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bcf7367809be234655459fb58416df18d2c934ff7aa01d722056f2c58952607:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d90b102cf05d661fc01da3ece406f1513f76ffb7f0bc495bb764402592c9e7c3:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:478b532e8faa47b25b9de0bee065feeb769b730b92ffc899a2e5480dbca53425:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1d6e3d1cf0cf8858c6abf2d11c5b0c2fc7dc15fbd42149561c30cd5135db91d2:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cd74dcea95243d581ced4a61484fc301a13e15ab030d5c1053aa01f81b721636:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1b355c69470a0de40c445d43f093b14be5ce3b7d706583e1e900f0152c444b28:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9604131a803386aaab76333069d918208607b2d0b6f45c54dfed1b46595aa80e:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b9756a8ee5ffcf994444c4a659269df981ff23fa7c4105f61e0a9fb6ca9654b1:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8935a2d3e658e3a7f77e99a719f086456fa57e7d702a17c63af10f9508fd5576:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:b5a2a77db1ce7dcd06457164e6ab0626e0b5111632ef0d0e88ff19aab6a978dd:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ed99cc03913a27e4f169b3451e10885aeee01070911ef7598a4efda5edb6756d:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b286645c7896efe2c1098d1e81a49b7b234ebd80d39bbf5934f13da395c3249:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ffcae266c66827c252488894e6181a1040d9a57892da90762390d5e356061012:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bad93c536031c4d79f68e5f491ebb7c5d4112ee99fafcfcc97bb6ed5e43fa9cd:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7bf17626fb841119cb77b140dcf340bb041c21adf336273c9815eb6e1dc0eed8:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:75c9e71352b71343cce7d9d2dde9e3ba56086081ba42e67bb2217f4cc0b4fc07:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fdc8c1703aadec811ac52bda8de50e3267061f8da931582e839664fe4a3a9a57:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:c6ace36e861ff0d56d6ce06dc19a43f8245b9097a312ea687685db7da09dde1c:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:07684a5ccf171bcaecc4270c8d81920cf42c38349d01e2409f11302e84546578:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1e6384198913c9f336971374bb1dd52c455c8f33d89853035e544aaf081fb384:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:45699c2ecd493e6b023d7b457521058494952eddc0bc52e36a6e97b16c014491:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:73a155285c9f87e6c232537491c21deb3a3e856826d5c9765e935aaf46609cdd:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e76495ab37317a5ec9ceee09bb3611a936d42ad09afe220fbb28f1dfc9d6ab1e:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:de0a9232afb9879ff954daecc26c22f843b4fe77240eb7eaf9ea281c8d67f482:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d06880e8db6e4925612d1e618167d4cecea8216c1ed0b1c636d005494046e08e:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:26f6ff99f7c6daf473a2ab4f651cf7fef31806aa024c41bbeba72366b93ed2fa:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:edb67ba78a9108fe0e20ad71108c29b17c5008a9e3d63a04b8208c51a77e14b2:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2fc2067c9cbf1b4b499561cc8780890f81906c121644bea7634c823dc516ab32:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:85df516e8fa4bb0192a27b8cfaa9cb77a6481ad3b7bc26412a2ce72cc334406e:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:41db1a58b629c7c003ee52c0941821b6c5e907685a4e3199eb937b38b0f36385:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6b7b0ad792ae795da40e271206aec6f8424b8afe0b501c16273aad4f8e3ae4fa:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:40f24a037044069ceb4f1e815669ce1e837c111851a3176c8daa7cf0a35143d9:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:51e1ec79336a48bd63819c225c7bfaf3085a5b07680e9ff1a84d7b7479ca8600:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ee44a3111bcce359516d7badf823a3aecacc00b01d2cc7ca79cc72c3414df2a9:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa80e2fd66b156d3795dd7733be9a735e837b6f21d65d8f75e7db9534c6504b5:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:d1dcd837c57222307f1bca7eec6629aab1f20dbe482b8babc11ecf67d7cad775:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f4316f439d5473f2d1d29983b3e536b76595226acc8b4de1027ae94d52f21ed:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fbe48b4d801687a6c4c3103cb76e030dd9a73b190beff8288155e8a78bedfa39:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6a8fa0cf870238c042072155dc12fb6a8af592c7e4c94cc751c6c0e9c65daa8a:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e225d40172060a0f4663d0766d4b895c8ad5b7ca70b412ec3cad10d284303647:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e801928b3a2902e6236e99e608a5a50f0b96a4b54e093edf636b64184019d854:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:661e803f639cea5453121397d48742d3075b5b9f9c8d25babe301f7ba0672c0d:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:df75560cdd5ce3fc639ba4f4b0f01e849bdb0d647775bba1bba4e536f876117d:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:63aa5e5c7f11cfc892172371fc0da4b7186e057f75a57688c86b451a03005295:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:clear_east", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:clear_north", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:clear_south", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7efd861570301596b2ecfa1bc9262640c7fced65ba2d96065d7fe1c460ee614c:clear_west", + "split": "calibration", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a943f6e5b7f8a6138198928b208eed5361d54d0fe18e6c524a9f78f49daa78c1:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 173, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:a6a6bea5658a948b1ac06450fc6179c496cb690fe076a3f1c118f3680bf4e895:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:91ed9b500eb8f32a013643bad9d7209049b9118d5b85b5f4953d784d8c6c09d8:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:fe7759d1cd794065852480e7b771a1d90907028acc454008ef503125f52c294f:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:20fa653670ee213486f4495f105d3ce1792965877283f3ab73b931ce002c5f52:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:67ab3d3c1ae00b14892340edfda91b3ec8bb507514a144bc42d7aaadfe5af201:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:ff63152c1fa9a41f12e9564906d36d395d7bfbdd6c42420c227e957a1a175d35:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:27acae5509e0822c228dffad9ad96b9aaec9054d976e279a00306f463d42fb7b:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7458e8fe0d014acfb2c6aa3c1ab90870c6084c96b27080dd67793c7ab697aefc:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0906eacbdc3a039fa3b5f2d7cee652f10c9ac2aa57a3bcd828e061f256a8405b:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2b7da97093f2d81401bdcb558eab993d03397e3fc96518474afbae13801c28b0:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:856a99db958ee53423aa2b5956b34a7b4f763f1487fdd8c41e09198833535db4:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cc16f99ea270270edce7ca4530b7b7f9abb1eff9412cdac3c36bb5aad7992ab8:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:850c888d8c14b99de5034d1ecd955ad7de19402697d6ae208c7251e976d217d8:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:10e94d5cad7405e31bbd0aa0ea376082c23aba6ea8d717d72e57873b589ae89d:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2edd03fd7639ba1505bd249db5828a1b06647879ca2862f00d3636baeae184ab:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f96efbd20d06a9b253650f80848aea15c4f26a25b3d661d4843c517604e386b2:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3fc5ea86c9c78c358f1fbb5fba48878966c04107fcfe721a6a9e07e3d47b2660:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4355aa990abb3ad5324fa524752e21b36133bf4c13ee3205954df81b7617bd1b:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 182, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bf3c33374b989324b4ea58b0b91d1381536f01a49ec8b51f049d22b6d0cda902:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:37da36de3adf8cedc3c30dc34df606eaa234a5bc75499b6390ee4f4908623559:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:7f6c7b6754ad744ad20dc574a746ea1383e4e8f8a86ca9819888aaa419e93a14:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:4f00b148927c3845ca34f8339d62d30a4cd5196658e4cd4ceedb371e7a74aa96:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:cae1d78e1545cfd1bd3152002c85a0fe14684bb73035fef3b11956182801ad80:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:823343d98311adef8f5565f78ef95e9e709802a94595385458dceb1192d2fa12:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 184, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 184, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 184, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8f352d54dfb7bb2b96c138c1a74c757d32365493a793034823be2e4a792bd9c2:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 184, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:150a3cdc45e1f68f39424146dbee23941aae7c4c429700eaeb719326086b240f:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f8a5d7beb108b5fef1ee7b68ad64b82710a97a89e8bf362891917eab72038c16:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8dcc705af2830ec1a7f0d09fa33b04d2ba96ff846e74f1fafb3c777d6cdf2b66:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2a38ca712fbac337fa87ad03e35a5f4dfe88e3865c92254c45368f40ff4fd20a:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:12d94be181b94f5d645ec61165329be25177907b3facb76ab5d45acd4f48469b:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f673ecc4571ea99a92a5e10b240c34497240c62d8f163e910e0acf94c0a830bc:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aa40833eb82d1d8b0d5bd65642032f77611702c817f0e64ddaa1110804c7a169:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:567795308aa69e6c6166dccbeb720c1499033d04cbe0ae8bc0483b6e742cc78f:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:f0b98046525ec488f013b857d4d0b2ba9413b11644881401cc4154f80d9f5e49:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:56603931fcc0d337599687e002bb24448b59286385fac5a28c31474efcfe6458:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:654a4894ae708735a9392667c5cc608daeb05a0a6cb661eaa21d0d998e43a7c5:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 185, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 185, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 185, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:aec26e9b85ab37603e8897f530b861b840ceace449149410a759f6ee7d4b0b68:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 185, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:66a103045db29551dfd9ec9ac9e3d7c5b7110ac1b6b53d1771a410b001da7399:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9687c8c103773f719c3e7777ea21b7e233b86f607e1c20b4b888b76fb3125dd4:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:8a40cf87b2684679a8f9b4d646aad34962d35acd403191b647301649347d7ac1:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:47919711a65150fa0fde78d1fb02c5c7d4b34a8b7c83e597cb86be095faf7f33:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:30bf36966cb06fa3d79f545b8abbc41d51de7e5a3b93b8aabcb8cc79d77f26ea:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 183, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:clear_east", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:clear_north", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:clear_south", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:1c7fd399b65e698175063ca0f813b1b41735384ae9ad3c6866256d6b9371170a:clear_west", + "split": "test", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0984da656b2437c647c52b18c127955ea24de75fb1f792aed30ec0a6fd485adb:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:42634df48d305619904375ed03b26f6f02ddd7b1e1dbe7720d358c9218a78244:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 172, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:963771d16e333ece9d42c14d74f1f20407a594f43f182d06bc81f8e1da0db091:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 181, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:23c176df28968b3d1c0028583b20ef971c2bc9074156775ba5e4bac19b2405d8:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:bd4106fc7afb8cfaea3cb8edb4a620bb84e44c8070d0edd18647615d1e9bec6d:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 175, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:335886524074e74138e2fc9693344bc0c9e73b0184d595430ab135730f6e5bcf:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:116f8e46a579a9314c000e5503e8b3469efaf44e27315f241e71e30e8d371759:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 176, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:9b7f547073ae6b84e390644e3cf1fdf2401882f30aa37a256ee2c8009ca7ddc3:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 180, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:3e4c972ddbf356a1380a1e569bbb7649c71be0fd418ddd18a03822a99a2886a2:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6ee3f501f86778b32c57917fb352832e08dc7bacad621356df267d568114ad57:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 174, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:e8073deaea4fea2347d37051dc8e411fb5f6982646e18421099fbb5bbfea754c:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:6cb16ec35df330032e48040cd32ba8effb8e15ea57c1c884f6c621cf0772d15c:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:5ee0de3e770e9d194358649333103c357bc048516cf5e93e60624bd516bf8b23:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:291173d7ba4e426cfeb8925acc7183a2359f09c0c6019e9386f2f5f58f3c69f5:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 179, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:2cfe879cc7f1dd5948d9c15c6e9bb9a8813a1e03cb5b08f84525e56d73cee720:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 178, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:clear_east", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:clear_north", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:clear_south", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + }, + { + "id": "maze_state:0cc19b0d392c1df73175282e246bebe15d8215d84b535aad657b04e52f194fe9:clear_west", + "split": "ood", + "family_id": "scaled_maze", + "type": "boolean", + "k": 2, + "max_path_tokens": 177, + "gold_distribution_kind": "deterministic_truth", + "gold_label_kind": "deterministic_truth", + "gold_distribution_usable": true, + "teacher_raw_sum": null, + "teacher_proxy_usable": false, + "teacher_argmax_matches_gold": null, + "target_transform": "quarantined_or_missing", + "teacher_target_error": null, + "teacher_raw_mapping": null + } +] diff --git a/variants/local_atomic_seed17/tokenizer/chat_template.jinja b/variants/local_atomic_seed17/tokenizer/chat_template.jinja new file mode 100644 index 0000000000000000000000000000000000000000..01be9b307daa2d425f7c168c9fb145a286e0afb4 --- /dev/null +++ b/variants/local_atomic_seed17/tokenizer/chat_template.jinja @@ -0,0 +1,89 @@ +{%- if tools %} + {{- '<|im_start|>system\n' }} + {%- if messages[0].role == 'system' %} + {{- messages[0].content + '\n\n' }} + {%- endif %} + {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within XML tags:\n" }} + {%- for tool in tools %} + {{- "\n" }} + {{- tool | tojson }} + {%- endfor %} + {{- "\n\n\nFor each function call, return a json object with function name and arguments within XML tags:\n\n{\"name\": , \"arguments\": }\n<|im_end|>\n" }} +{%- else %} + {%- if messages[0].role == 'system' %} + {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }} + {%- endif %} +{%- endif %} +{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %} +{%- for message in messages[::-1] %} + {%- set index = (messages|length - 1) - loop.index0 %} + {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('') and message.content.endswith('')) %} + {%- set ns.multi_step_tool = false %} + {%- set ns.last_query_index = index %} + {%- endif %} +{%- endfor %} +{%- for message in messages %} + {%- if message.content is string %} + {%- set content = message.content %} + {%- else %} + {%- set content = '' %} + {%- endif %} + {%- if (message.role == "user") or (message.role == "system" and not loop.first) %} + {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }} + {%- elif message.role == "assistant" %} + {%- set reasoning_content = '' %} + {%- if message.reasoning_content is string %} + {%- set reasoning_content = message.reasoning_content %} + {%- else %} + {%- if '' in content %} + {%- set reasoning_content = content.split('')[0].rstrip('\n').split('')[-1].lstrip('\n') %} + {%- set content = content.split('')[-1].lstrip('\n') %} + {%- endif %} + {%- endif %} + {%- if loop.index0 > ns.last_query_index %} + {%- if loop.last or (not loop.last and reasoning_content) %} + {{- '<|im_start|>' + message.role + '\n\n' + reasoning_content.strip('\n') + '\n\n\n' + content.lstrip('\n') }} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- else %} + {{- '<|im_start|>' + message.role + '\n' + content }} + {%- endif %} + {%- if message.tool_calls %} + {%- for tool_call in message.tool_calls %} + {%- if (loop.first and content) or (not loop.first) %} + {{- '\n' }} + {%- endif %} + {%- if tool_call.function %} + {%- set tool_call = tool_call.function %} + {%- endif %} + {{- '\n{"name": "' }} + {{- tool_call.name }} + {{- '", "arguments": ' }} + {%- if tool_call.arguments is string %} + {{- tool_call.arguments }} + {%- else %} + {{- tool_call.arguments | tojson }} + {%- endif %} + {{- '}\n' }} + {%- endfor %} + {%- endif %} + {{- '<|im_end|>\n' }} + {%- elif message.role == "tool" %} + {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %} + {{- '<|im_start|>user' }} + {%- endif %} + {{- '\n\n' }} + {{- content }} + {{- '\n' }} + {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %} + {{- '<|im_end|>\n' }} + {%- endif %} + {%- endif %} +{%- endfor %} +{%- if add_generation_prompt %} + {{- '<|im_start|>assistant\n' }} + {%- if enable_thinking is defined and enable_thinking is false %} + {{- '\n\n\n\n' }} + {%- endif %} +{%- endif %} \ No newline at end of file diff --git a/variants/local_atomic_seed17/tokenizer/tokenizer.json b/variants/local_atomic_seed17/tokenizer/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..c7afbed2efcdf019f88ab0572ec29d3bf595dfe2 --- /dev/null +++ b/variants/local_atomic_seed17/tokenizer/tokenizer.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506 +size 11422650 diff --git a/variants/local_atomic_seed17/tokenizer/tokenizer_config.json b/variants/local_atomic_seed17/tokenizer/tokenizer_config.json new file mode 100644 index 0000000000000000000000000000000000000000..9a87e3822403fb40a9524bb75edb66c6f0585750 --- /dev/null +++ b/variants/local_atomic_seed17/tokenizer/tokenizer_config.json @@ -0,0 +1,30 @@ +{ + "add_prefix_space": false, + "backend": "tokenizers", + "bos_token": null, + "clean_up_tokenization_spaces": false, + "eos_token": "<|im_end|>", + "errors": "replace", + "extra_special_tokens": [ + "<|im_start|>", + "<|im_end|>", + "<|object_ref_start|>", + "<|object_ref_end|>", + "<|box_start|>", + "<|box_end|>", + "<|quad_start|>", + "<|quad_end|>", + "<|vision_start|>", + "<|vision_end|>", + "<|vision_pad|>", + "<|image_pad|>", + "<|video_pad|>" + ], + "is_local": true, + "local_files_only": true, + "model_max_length": 131072, + "pad_token": "<|endoftext|>", + "split_special_tokens": false, + "tokenizer_class": "Qwen2Tokenizer", + "unk_token": null +} diff --git a/variants/local_atomic_seed17/train_log.json b/variants/local_atomic_seed17/train_log.json new file mode 100644 index 0000000000000000000000000000000000000000..1d391a14198a7f7d4978c1d1b065c9414a5ab7b2 --- /dev/null +++ b/variants/local_atomic_seed17/train_log.json @@ -0,0 +1,2924 @@ +[ + { + "step": 1, + "phase": "full", + "loss": 0.7357752248644829, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 1.1536121107637882, + "batch_question_ids_sha256": "bd278a5381dfd04522805f010f2dc74e31169ee69bf85f4bf798a450e6a5e3f7" + }, + { + "step": 2, + "phase": "full", + "loss": 0.761392205953598, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 1.9856814863160253, + "batch_question_ids_sha256": "0fe5136dace4ce8479cf66de275ce14b846ddb0c19f0f74981f1218b69f8b721" + }, + { + "step": 3, + "phase": "full", + "loss": 0.7301052659749985, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 2.692872053012252, + "batch_question_ids_sha256": "926a2cee8dd2138ad485d3ca6fb573c5fd910d66c1ae1d13add77ca9e0a1d4d0" + }, + { + "step": 4, + "phase": "full", + "loss": 0.7156264334917068, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 3.3986059725284576, + "batch_question_ids_sha256": "705a290c921844619149a334b7d81601595b2f2fad01ded5750bdf2ba40c9653" + }, + { + "step": 5, + "phase": "full", + "loss": 0.7088156193494797, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 4.106619888916612, + "batch_question_ids_sha256": "5b7662fb22d7a4a2596c520982e1ceb13edb394e1b2eab1ef74877bf8cdab9a0" + }, + { + "step": 6, + "phase": "full", + "loss": 0.5955959856510162, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 4.821541517972946, + "batch_question_ids_sha256": "d750f325bc4c4efc0981e540bfccaf9faaf85e6a93c9d3d03ea16201f15ea95f" + }, + { + "step": 7, + "phase": "full", + "loss": 0.7145266979932785, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 5.532915022224188, + "batch_question_ids_sha256": "e423e75f65653bd2883b4ddc4aefdec23a98dfc64981963b28598b858f6ad1ff" + }, + { + "step": 8, + "phase": "full", + "loss": 0.6962291747331619, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 6.241126445122063, + "batch_question_ids_sha256": "ed2caf2b8f1ce1b0167988400ac6913f646c98889d4f9af8ec556b01d296098e" + }, + { + "step": 9, + "phase": "full", + "loss": 0.6686204820871353, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 6.95830041822046, + "batch_question_ids_sha256": "97dc8ec4450263617bd08b6bcd91a13ed6f0caa36fd59f18d1a0b8b55dee27d4" + }, + { + "step": 10, + "phase": "full", + "loss": 0.6946953684091568, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 7.6704751485958695, + "batch_question_ids_sha256": "3bf6153848b180200f761236e990d6fa540a88546a9316419179d7e8e050b586" + }, + { + "step": 11, + "phase": "full", + "loss": 0.6810178160667419, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 8.379737825132906, + "batch_question_ids_sha256": "0421d6e3782e74b84a56a946b47814c4f1c5ddf000bef5c4f26efb206ee6413b" + }, + { + "step": 12, + "phase": "full", + "loss": 0.7251933217048645, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 9.085399012081325, + "batch_question_ids_sha256": "0c859c501ee768c0f13ab06798a8dae5f5175e9df6a7617865afde1b082f4260" + }, + { + "step": 13, + "phase": "full", + "loss": 0.6593339741230011, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 9.797674008645117, + "batch_question_ids_sha256": "124d929a8b02b6bef8c8d504f573a10bc7e86cd45f5a44929548064b3c2e4a2b" + }, + { + "step": 14, + "phase": "full", + "loss": 0.6707886457443237, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 10.52715971879661, + "batch_question_ids_sha256": "8f668a5c5e031df68dfb6c0c540cee8fa7fd1e004a13a401a0ccd16ee6ea0891" + }, + { + "step": 15, + "phase": "full", + "loss": 0.725605234503746, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 11.336303167045116, + "batch_question_ids_sha256": "267a2eea408cd6f51265862fe6cddb92915a3bd6deeae5e1284ba364ad86556b" + }, + { + "step": 16, + "phase": "full", + "loss": 0.7417129427194595, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 12.044057726860046, + "batch_question_ids_sha256": "4679ccb6c511a07cb0fdfee426f2efcec64249f3e15685b17780a37e308278fc" + }, + { + "step": 17, + "phase": "full", + "loss": 0.6930674612522125, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 12.764025431126356, + "batch_question_ids_sha256": "36923a26d90a0a300d07fd1c880628072165547a8216a825b9e6860984c69d17" + }, + { + "step": 18, + "phase": "full", + "loss": 0.6594993025064468, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 13.469019323587418, + "batch_question_ids_sha256": "2959b9ecf5aba65dc6c64bbb31c05f56c6ea774ba1c94df624c72ceb34db3936" + }, + { + "step": 19, + "phase": "full", + "loss": 0.6686153262853622, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 14.180168819613755, + "batch_question_ids_sha256": "8020d8e6e9e9e45fbb56809f9b0a73e7e2fe5c8794dad738cdb0fc5a67174dc0" + }, + { + "step": 20, + "phase": "full", + "loss": 0.692922979593277, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 14.885747655294836, + "batch_question_ids_sha256": "35d8d0ed4c834a581709cca72ed37656ea37d206c1e9230cfe9cc016581f2a00" + }, + { + "step": 21, + "phase": "full", + "loss": 0.6937933564186096, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 15.614209344610572, + "batch_question_ids_sha256": "0a49456a8e1ca970c08d00458ff1176df57dfe40ba9fb9ae89cc9dc9cdd06c36" + }, + { + "step": 22, + "phase": "full", + "loss": 0.6798531860113144, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 16.317013545893133, + "batch_question_ids_sha256": "d3d28375c2bc4499a85c16e3716523e4ada50e46b0f540a58dac00f1fbf417d6" + }, + { + "step": 23, + "phase": "full", + "loss": 0.6879730224609375, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 17.023230530321598, + "batch_question_ids_sha256": "69d3fd53a9ea304e7f7978f4b8092a33ef4989b74a52944473a25f170360a880" + }, + { + "step": 24, + "phase": "full", + "loss": 0.677180677652359, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 17.726934481412172, + "batch_question_ids_sha256": "612206abce97fc7417f161402a63024fa3d15f014bb97f5e783b50cdaed1f0c7" + }, + { + "step": 25, + "phase": "full", + "loss": 0.7006799876689911, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 18.432671025395393, + "batch_question_ids_sha256": "04c152b2f198c3a6c26606a59c6d171c866d43b9df1dc3f555da775a0717d882" + }, + { + "step": 26, + "phase": "full", + "loss": 0.6874385625123978, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 19.13509521447122, + "batch_question_ids_sha256": "2083d5f2f408a840ab8c8db3adcb101956434a3d95fa7f1332dcec35c01fc22d" + }, + { + "step": 27, + "phase": "full", + "loss": 0.6986180990934372, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 19.86363977473229, + "batch_question_ids_sha256": "829de0de538bf35f619918408bc91027f07602170e00583d67fc680d316deb72" + }, + { + "step": 28, + "phase": "full", + "loss": 0.6676506698131561, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 20.63883249554783, + "batch_question_ids_sha256": "04baaab0326d45beb11ab0620abff7f1f880aef21734a8d3ae04bd888844086a" + }, + { + "step": 29, + "phase": "full", + "loss": 0.6577211171388626, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 21.356288520619273, + "batch_question_ids_sha256": "9bf5e69fd002a1dc6ece3698c14caf5260e417831c205490f66f2906d14fcfb1" + }, + { + "step": 30, + "phase": "full", + "loss": 0.7583383917808533, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 22.058414204977453, + "batch_question_ids_sha256": "84638dde2892cd17608cacbb8af0edc68a2494855a1549d977afe0a327b5a2a5" + }, + { + "step": 31, + "phase": "full", + "loss": 0.6728695780038834, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 22.76617749221623, + "batch_question_ids_sha256": "6a70bfe868d73d5a56aaba10484706e440d8f5c1538ffd60d2e9c0724ea769ff" + }, + { + "step": 32, + "phase": "full", + "loss": 0.8057665228843689, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 23.468083929270506, + "batch_question_ids_sha256": "aded211d908d79ff3011b780d1367a46b2da083cb6aa18c1b3283bb50009da56" + }, + { + "step": 33, + "phase": "full", + "loss": 0.6377183496952057, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 24.176365829072893, + "batch_question_ids_sha256": "ae09c436570a828ae039864afa81e99c5a0f16998bf180e8aff63ec9cb38241b" + }, + { + "step": 34, + "phase": "full", + "loss": 0.6877293884754181, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 24.880596259608865, + "batch_question_ids_sha256": "97c0ca0d50666eabdd0f774886e3c6fef2ca6baf5a0c4fb745ce01b98e2b0858" + }, + { + "step": 35, + "phase": "full", + "loss": 0.6296602934598923, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 25.588200164027512, + "batch_question_ids_sha256": "82c1758fec76db40a5bd577abc58f5ce6bb008a2365aea984bf5a8db6e451b8c" + }, + { + "step": 36, + "phase": "full", + "loss": 0.6966435015201569, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 26.29398136585951, + "batch_question_ids_sha256": "1b4d887afd458bcf93718c769d6279c8a87214d723f621eb296e11fe76b1e413" + }, + { + "step": 37, + "phase": "full", + "loss": 0.6992466896772385, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 27.002122561447322, + "batch_question_ids_sha256": "914569245a9a0183d982cc00bedf25f551fdef5bb6f04cb2a62ef873246840c6" + }, + { + "step": 38, + "phase": "full", + "loss": 0.6085568517446518, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 27.709516515024006, + "batch_question_ids_sha256": "6620d8eb184ab8f884b95152752f6b8bb08f0e8f874ba210e7389da581322c3e" + }, + { + "step": 39, + "phase": "full", + "loss": 0.6955265253782272, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 28.424468234181404, + "batch_question_ids_sha256": "1fa06b4ecdc5772e671e9985bba6058d8f101a60a7774d46f8736ef4b6a046cc" + }, + { + "step": 40, + "phase": "full", + "loss": 0.7289998978376389, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 29.1672170907259, + "batch_question_ids_sha256": "b1201c2810defae43cd07ac84609ead35146bc4b4ffc1cafd5ac08f16e08bc1a" + }, + { + "step": 41, + "phase": "full", + "loss": 0.7442513257265091, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 29.965257606469095, + "batch_question_ids_sha256": "d6defc8ab30a49b36dfd6f31a1c6934cf5f0d407b8d31726ac589682b89869f0" + }, + { + "step": 42, + "phase": "full", + "loss": 0.6680125445127487, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 30.671857598237693, + "batch_question_ids_sha256": "509d08bad6c2dbd890a3abce4d6fa0a44a2291886dd0c95cf5f3d9464b6e2b10" + }, + { + "step": 43, + "phase": "full", + "loss": 0.7109745144844055, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 31.38111410290003, + "batch_question_ids_sha256": "14ac55d2d4bfc77ba35fabf6177f7b7c42c42a1ea4e34805eee8837c8d821b1e" + }, + { + "step": 44, + "phase": "full", + "loss": 0.6921908259391785, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 32.0849255444482, + "batch_question_ids_sha256": "019d50b0f6da17de22bfa21e6528f7468027f3afa647cd2bf4ec5168a97cec8f" + }, + { + "step": 45, + "phase": "full", + "loss": 0.6866155415773392, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 32.79270731750876, + "batch_question_ids_sha256": "5bce9a477d3827053043e7ea02d34fad2cc955a750be5b76b4ab55e5c4d000bd" + }, + { + "step": 46, + "phase": "full", + "loss": 0.7431487888097763, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 33.49928600806743, + "batch_question_ids_sha256": "20e21fde428716685c0db1942ac7930f9848578be8006d08e60b042f08a1ba77" + }, + { + "step": 47, + "phase": "full", + "loss": 0.7528977543115616, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 34.209237270988524, + "batch_question_ids_sha256": "a91eb064de49c72915c9b37bcf3b92744a404421cf9e24753dbc964733ba260b" + }, + { + "step": 48, + "phase": "full", + "loss": 0.7595830708742142, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 34.91343867499381, + "batch_question_ids_sha256": "a1a27410b35631de3bcd298f73af27812c62e5ec7c50e1f7be0400627cdf769b" + }, + { + "step": 49, + "phase": "full", + "loss": 0.6752226948738098, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 35.62406632024795, + "batch_question_ids_sha256": "235daa1b6aed7dc5e0c7454df395923d8d7c769da58d134d8615cce30988d4b6" + }, + { + "step": 50, + "phase": "full", + "loss": 0.6898482739925385, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 36.333745242096484, + "batch_question_ids_sha256": "4ed5ac62a5c9eba49467b524168d9f1973dc28d091e62331615d995df7f93d22", + "dev": { + "questions": 192, + "objective": "gold_distribution", + "eligible_questions": 192, + "target_ce": 0.6719256625260513, + "target_kl": 0.6719256625260513, + "target_tv": 0.4873879654301854, + "by_target_kind": { + "deterministic_truth": { + "n": 192, + "ce": 0.6719256625260513, + "kl": 0.6719256625260513, + "tv": 0.4873879654301854 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 192, + "ce": 0.6719256625260513, + "kl": 0.6719256625260513, + "tv": 0.4873879654301854 + }, + "observed_outcome": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + { + "step": 51, + "phase": "full", + "loss": 0.69352225959301, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 41.09118528198451, + "batch_question_ids_sha256": "22378d3557d9898076f2cd531c1bbed491a290f5a0b8879b2ae5c3f0751f5144" + }, + { + "step": 52, + "phase": "full", + "loss": 0.6345449686050415, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 41.8039462082088, + "batch_question_ids_sha256": "077d687a1662bbd6fe30d5c55ddfc31a660210f622b90f42227c1212749c1e44" + }, + { + "step": 53, + "phase": "full", + "loss": 0.7950070351362228, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 42.630774778313935, + "batch_question_ids_sha256": "8075963f53ca9f441e87ec0e1b579eaeb7c12610e90c3bc704528efbc1d68caa" + }, + { + "step": 54, + "phase": "full", + "loss": 0.7242231965065002, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 43.33911581430584, + "batch_question_ids_sha256": "77bcfb05a83aa9fcf8728f4e21f297cf99a044721762e9941ec524934bb80931" + }, + { + "step": 55, + "phase": "full", + "loss": 0.7274714410305023, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 44.04821956157684, + "batch_question_ids_sha256": "e80edb5ee0ef09558f073c86b6222c57115fee8469e76738bfe8f1b96e1c9a11" + }, + { + "step": 56, + "phase": "full", + "loss": 0.7674485445022583, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 44.75327822007239, + "batch_question_ids_sha256": "230503d10db92c92e906238b0b9e678ee0a3c6dd6316db01a9a81a1b7fc5cbca" + }, + { + "step": 57, + "phase": "full", + "loss": 0.7148261070251465, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 45.47067920677364, + "batch_question_ids_sha256": "d3655962682af8cb5719aeb7b5777826e71c3c2788f7017503cfb165441dc068" + }, + { + "step": 58, + "phase": "full", + "loss": 0.7253440022468567, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 46.17538301181048, + "batch_question_ids_sha256": "08e4cf2b309703ac4d91b9cdf4c3a46dcb086963afe07ef4f4dc61e7c5b63b73" + }, + { + "step": 59, + "phase": "full", + "loss": 0.6369745433330536, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 46.88810151536018, + "batch_question_ids_sha256": "883d828d3e197bdac7f279c14cd9dda5ebb0efdc046459287b9f582882488b82" + }, + { + "step": 60, + "phase": "full", + "loss": 0.6272382587194443, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 47.59516160748899, + "batch_question_ids_sha256": "d3afb63e3075e15c65ec94563a2f3ebeba7f895e52f7c61d71a85543b984f565" + }, + { + "step": 61, + "phase": "full", + "loss": 0.6565569788217545, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 48.30792665109038, + "batch_question_ids_sha256": "6be56e055f0fcd2091aea55acd23ff88f88b07825144f375ec3bd5bee966e90d" + }, + { + "step": 62, + "phase": "full", + "loss": 0.6718819588422775, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 49.01261640898883, + "batch_question_ids_sha256": "a745e4e9a7848c58d1888bb8d349d8a62a13a2dcfecc3e08dbbab789f8e0a92b" + }, + { + "step": 63, + "phase": "full", + "loss": 0.6767185479402542, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 49.71684215031564, + "batch_question_ids_sha256": "09299f45d10be3d63db8f17fd32f525fcd2ddd6b74e2d2c05ee1c2d691548661" + }, + { + "step": 64, + "phase": "full", + "loss": 0.7076004594564438, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 50.42292621452361, + "batch_question_ids_sha256": "4b1a479368b0d1cd481d2b68e0781e0ce802d461df9937e6b4f26caa781611fa" + }, + { + "step": 65, + "phase": "full", + "loss": 0.6769139021635056, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 51.14412056654692, + "batch_question_ids_sha256": "61b5c22be18edfe8d99d4e7c9a4cc5fdc0a2b9afb9bf6290aa44779793a40e8b" + }, + { + "step": 66, + "phase": "full", + "loss": 0.6360280811786652, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 51.940813385881484, + "batch_question_ids_sha256": "3a30969db79423d47e0e7f365ecec4e58799345cb23ca61c9ee59faaf2479157" + }, + { + "step": 67, + "phase": "full", + "loss": 0.6980292052030563, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 52.65448736585677, + "batch_question_ids_sha256": "36d2aa92af1a1628b9caecdec0fe07738182b6eae750faa51b3a979e37eef5c7" + }, + { + "step": 68, + "phase": "full", + "loss": 0.6933810859918594, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 53.36218072660267, + "batch_question_ids_sha256": "e48cbf39e56fbf7a22f03e8f1c6fbec5088b892eacad96b2a144e617839bf373" + }, + { + "step": 69, + "phase": "full", + "loss": 0.6672176718711853, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 54.07358059100807, + "batch_question_ids_sha256": "1f42a07987351a4ae9e8f2d960f138237acba01a57f7c4d168981b7385399333" + }, + { + "step": 70, + "phase": "full", + "loss": 0.6666816920042038, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 54.786875791847706, + "batch_question_ids_sha256": "dc732fc291bda20338e365684f57e8013a61ece5d7e9e51044a1f34c4209b8cf" + }, + { + "step": 71, + "phase": "full", + "loss": 0.6898149847984314, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 55.496858866885304, + "batch_question_ids_sha256": "feaf0325f226102c7068e352d946be539bd6c53f500fefd27e122ed08091643d" + }, + { + "step": 72, + "phase": "full", + "loss": 0.69078728556633, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 56.20261433906853, + "batch_question_ids_sha256": "faff1f92980310c4577eb20a6f3f8df705bf0afc093ed29a17e5468e91648247" + }, + { + "step": 73, + "phase": "full", + "loss": 0.6899441331624985, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 56.91690556611866, + "batch_question_ids_sha256": "868a9be578a23b6efad51e8daef751d7a3db0ec1769d5fcc66d2c6f44cb8db0f" + }, + { + "step": 74, + "phase": "full", + "loss": 0.6904614120721817, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 57.627597698010504, + "batch_question_ids_sha256": "aef64813c8cad9b76e9827e4f060263acc426e265d1f147a72a18a406e2e1dec" + }, + { + "step": 75, + "phase": "full", + "loss": 0.6478376388549805, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 58.34027614351362, + "batch_question_ids_sha256": "87b75a201e60f6c2c720c85875a3ad932b571fd36dd93472091eac790bfacfba" + }, + { + "step": 76, + "phase": "full", + "loss": 0.6931174099445343, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 59.04812639951706, + "batch_question_ids_sha256": "59c48438ae3db56855b95e15e339552947a8aa1c997ec78e790cbb1b26bb9391" + }, + { + "step": 77, + "phase": "full", + "loss": 0.6702842265367508, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 59.75835583731532, + "batch_question_ids_sha256": "03d201ed418265779a2c4ab393c3c9e4058a21094bc18d35d4b620403b00735b" + }, + { + "step": 78, + "phase": "full", + "loss": 0.666019931435585, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 60.471717577427626, + "batch_question_ids_sha256": "3c81b9499d126d056c435dfcb9ccaa24d1e0767627dfe8a40da14b2f0d45d332" + }, + { + "step": 79, + "phase": "full", + "loss": 0.6347012966871262, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 61.246109472587705, + "batch_question_ids_sha256": "c159b5f98a2e64919d48d001462ae379233fd3a467572f6bc79a96ef23f1ad9e" + }, + { + "step": 80, + "phase": "full", + "loss": 0.7028255760669708, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 61.951375235803425, + "batch_question_ids_sha256": "88df0b463d0172becbc5e2532438dd55da1223c604ffcbf7dd6f14e69810fa97" + }, + { + "step": 81, + "phase": "full", + "loss": 0.6976883262395859, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 62.663546139374375, + "batch_question_ids_sha256": "59b27a09f0675ab80ee79ef454a897ea318276dfe9ddf022d366a46f9592195c" + }, + { + "step": 82, + "phase": "full", + "loss": 0.7760961502790451, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 63.37679443694651, + "batch_question_ids_sha256": "f730ad7be0c65016e8b5c74d830ded95d519c0f564a0f3e828ab1c5d485ea04e" + }, + { + "step": 83, + "phase": "full", + "loss": 0.7888075858354568, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 64.08315645437688, + "batch_question_ids_sha256": "ed147ffe9afdb7ac411d8098acf769dff0621408cbd12581109a1c9413632459" + }, + { + "step": 84, + "phase": "full", + "loss": 0.7171074450016022, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 64.79158241581172, + "batch_question_ids_sha256": "499e97ed95fb67ee457df92fe98472902730bee530dab26de608335068a6731a" + }, + { + "step": 85, + "phase": "full", + "loss": 0.6482377052307129, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 65.49712947010994, + "batch_question_ids_sha256": "37cfeaab1602d95710b3e233afe447bffe919edac977cb1ee1ac80b902171030" + }, + { + "step": 86, + "phase": "full", + "loss": 0.6591252312064171, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 66.21679510455579, + "batch_question_ids_sha256": "64761283bc5b9e305afdb15be2a678ca74832a7e7f081b2ec4b6aab09fbc15c6" + }, + { + "step": 87, + "phase": "full", + "loss": 0.5492922067642212, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 66.92293965630233, + "batch_question_ids_sha256": "e708b782fa5105fb7fc5cd45122d342c9eeb75fadcf95e08497c3c37d97e2059" + }, + { + "step": 88, + "phase": "full", + "loss": 0.6606456190347672, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 67.63652727846056, + "batch_question_ids_sha256": "ab10f35b94a604a98bdf4a84b59011f0508f658eb0f54aa15575ab35abda482c" + }, + { + "step": 89, + "phase": "full", + "loss": 0.6908959448337555, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 68.34350839536637, + "batch_question_ids_sha256": "9ae77408aa58a26a2c3cb57e215b2f75868621633a5af68bc298017fb2b9a09a" + }, + { + "step": 90, + "phase": "full", + "loss": 0.6743407547473907, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 69.05191356595606, + "batch_question_ids_sha256": "0bbec5efbded727ca587719c93e0c0bd530f68f511ebe822a8ef2c3c0179630c" + }, + { + "step": 91, + "phase": "full", + "loss": 0.6646024733781815, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 69.76550601795316, + "batch_question_ids_sha256": "a25ad13539f325a629886b6efac3d070deb7dbeb0f4af8e2afb4a89303ca9259" + }, + { + "step": 92, + "phase": "full", + "loss": 0.6374963521957397, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 70.57190259732306, + "batch_question_ids_sha256": "2e9940ff9bf0925f020a47fff5ba0c39871c2cb136576d4465ba154a3e043f6c" + }, + { + "step": 93, + "phase": "full", + "loss": 0.6388277858495712, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 71.27765150554478, + "batch_question_ids_sha256": "1ebf8eb5b19bebd8068017545e450f5779e0a2f138ddc2c7e0651a7d410fe119" + }, + { + "step": 94, + "phase": "full", + "loss": 0.7133028507232666, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 71.98841159045696, + "batch_question_ids_sha256": "2ef585b3ea019c00474160f7028592079a93b1493a63054a787a3d37be087132" + }, + { + "step": 95, + "phase": "full", + "loss": 0.8345339745283127, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 72.70448725577444, + "batch_question_ids_sha256": "8dc517693815691fe2fc39dd0f1137e116b3f3f7b90fbb8c73036da89474727f" + }, + { + "step": 96, + "phase": "full", + "loss": 0.704349011182785, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 73.41507052537054, + "batch_question_ids_sha256": "eb80fc706e4f4add3a28538f73b6b92de73d22b10320ab9daac5da25bfab85a5" + }, + { + "step": 97, + "phase": "full", + "loss": 0.6171626895666122, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 74.12457055598497, + "batch_question_ids_sha256": "41dd2777fa3dc0b40f1da4e23a8fa9f8ac7fe1d417c7c25c8ba822990db5d97f" + }, + { + "step": 98, + "phase": "full", + "loss": 0.6930581629276276, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 74.83647784404457, + "batch_question_ids_sha256": "93f015a0e797dcb0c924a26329bd301fa561831babb50ccc188990398c87af53" + }, + { + "step": 99, + "phase": "full", + "loss": 0.7146369367837906, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 75.54242298658937, + "batch_question_ids_sha256": "1278124e3b3b7ed48c8a0093b7adb24e21a901ad2ab8cc2b0034938eac9ac67b" + }, + { + "step": 100, + "phase": "full", + "loss": 0.6825987249612808, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 76.25459717493504, + "batch_question_ids_sha256": "38e850934e49b0861f753f2d78a1b6aab3af7a8ac410f4c63e9fe6a249f78231", + "dev": { + "questions": 192, + "objective": "gold_distribution", + "eligible_questions": 192, + "target_ce": 0.6561035135692814, + "target_kl": 0.6561035135692814, + "target_tv": 0.46386252848605275, + "by_target_kind": { + "deterministic_truth": { + "n": 192, + "ce": 0.6561035135692814, + "kl": 0.6561035135692814, + "tv": 0.46386252848605275 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 192, + "ce": 0.6561035135692814, + "kl": 0.6561035135692814, + "tv": 0.46386252848605275 + }, + "observed_outcome": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + { + "step": 101, + "phase": "full", + "loss": 0.6906978040933609, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 81.69229925051332, + "batch_question_ids_sha256": "dde8c9a4c60ec78625029c019f228230243d7482ec609fa0176b293927b184b2" + }, + { + "step": 102, + "phase": "full", + "loss": 0.6857090294361115, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 82.4013346945867, + "batch_question_ids_sha256": "617db5dbbb8fa7d8f507e98ac833c2d115784b8f7ec1fb65960cc8e4e18c7d6d" + }, + { + "step": 103, + "phase": "full", + "loss": 0.7056168615818024, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 83.11017606779933, + "batch_question_ids_sha256": "55f1a7964395e6c6efdc7b9d7befd2e1bb81423e0b0bdba27f6686b1d72ea821" + }, + { + "step": 104, + "phase": "full", + "loss": 0.6894330531358719, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 83.82529752794653, + "batch_question_ids_sha256": "8eec335b9ec233ad882a253b6c4776b5b38d807cb6057db8c7991e50ae41ef72" + }, + { + "step": 105, + "phase": "full", + "loss": 0.5372484773397446, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 84.61041427776217, + "batch_question_ids_sha256": "887bf7f9792a3a5fc1e96033b43856da47902f70c9eed7c779bb2fb4a59ffffe" + }, + { + "step": 106, + "phase": "full", + "loss": 0.6504513472318649, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 85.31767488177866, + "batch_question_ids_sha256": "1114a914b4e89c3cb9df8484e14529e0aa574dbe7c7f5885f50c1eb657f6f74e" + }, + { + "step": 107, + "phase": "full", + "loss": 0.6374363452196121, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 86.02894496452063, + "batch_question_ids_sha256": "6bcbeb41d505302819847fe9b2dc17e4027ead8f8c34371486af8ed29927e0dd" + }, + { + "step": 108, + "phase": "full", + "loss": 0.6749598979949951, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 86.7358630578965, + "batch_question_ids_sha256": "0268d55a1dc9dd2b151f417884195395eda83ed4948a7161b0584941e9e8cea0" + }, + { + "step": 109, + "phase": "full", + "loss": 0.6306131780147552, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 87.44477162323892, + "batch_question_ids_sha256": "0b9b6fda91ddcb27a4dff9b765587b9fc416058afa07ef6fc36dd67521e0bc82" + }, + { + "step": 110, + "phase": "full", + "loss": 0.6189803555607796, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 88.14843521546572, + "batch_question_ids_sha256": "d345f86330c28392e173b75bcdd19b605605b20d49426b3ac479ba99e0af3aeb" + }, + { + "step": 111, + "phase": "full", + "loss": 0.7491437494754791, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 88.86078819446266, + "batch_question_ids_sha256": "e62f0e69301b683a423ac0d90903ac026962c3e8a950d8feccf8be7ff09db7c5" + }, + { + "step": 112, + "phase": "full", + "loss": 0.6661238819360733, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 89.55585942883044, + "batch_question_ids_sha256": "d642597a2fe0c63b7cec8375b8dfc33a96a584368dfe1a4dd79869814a3c1b25" + }, + { + "step": 113, + "phase": "full", + "loss": 0.545436218380928, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 90.26204445026815, + "batch_question_ids_sha256": "0e54051a97c65ed53e6e3b027267f445bef288e5641566746c11f12001becfb1" + }, + { + "step": 114, + "phase": "full", + "loss": 0.7130173891782761, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 90.96647855918854, + "batch_question_ids_sha256": "76b0864c8f26baa5afcc103d26ab7396ceb2e60a5393af389d6e7cc89b2ca36f" + }, + { + "step": 115, + "phase": "full", + "loss": 0.605145201086998, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 91.674177730456, + "batch_question_ids_sha256": "9d936cd3408556568a914131cd54b722f67f7ea742a65aab051ab2cbe40a0a02" + }, + { + "step": 116, + "phase": "full", + "loss": 0.5762528702616692, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 92.37963872496039, + "batch_question_ids_sha256": "a0f9d5fec47dde94d2cdd5fd1b9eb97b246e08a85a89ba81e9cd7c64518534f6" + }, + { + "step": 117, + "phase": "full", + "loss": 0.6933784335851669, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 93.10190506279469, + "batch_question_ids_sha256": "16d7545fbd452f989106681af726f7284e128d3975845e07af227ed3c71b5964" + }, + { + "step": 118, + "phase": "full", + "loss": 0.7249573320150375, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 93.88858453836292, + "batch_question_ids_sha256": "762da79ffa5005b17ebf331bcf6cc7963cde4e78740cfa9438d8082c0fec3af0" + }, + { + "step": 119, + "phase": "full", + "loss": 0.7451318651437759, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 94.59799578599632, + "batch_question_ids_sha256": "b1779d151bf71bec852b5507245bbb171c3c7aa8b08b046bf7ae048bc270d978" + }, + { + "step": 120, + "phase": "full", + "loss": 0.7655170857906342, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 95.30078157223761, + "batch_question_ids_sha256": "758af5f9695609216c7f784070502ffe55f7d5c11cda473b7ee1e9d85668ed6e" + }, + { + "step": 121, + "phase": "full", + "loss": 0.7155108898878098, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 96.00860889349133, + "batch_question_ids_sha256": "fbcc14e82236c412265f000b8bdefc961c10e87eb2b73393392342d080b83217" + }, + { + "step": 122, + "phase": "full", + "loss": 0.7464017271995544, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 96.71244681626558, + "batch_question_ids_sha256": "5bcec80c67f2808394255477064542eab870d99b6a72e89e4d7079a2876a4825" + }, + { + "step": 123, + "phase": "full", + "loss": 0.6493610143661499, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 97.41965732257813, + "batch_question_ids_sha256": "218dd2d5fdae414da4888ada98956114b4caaca1d66d952d76680189edf32b43" + }, + { + "step": 124, + "phase": "full", + "loss": 0.7181983590126038, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 98.12668322678655, + "batch_question_ids_sha256": "4d16681b462c3a0e44403f30e6095681f5a959721483166fea71921a2a80350b" + }, + { + "step": 125, + "phase": "full", + "loss": 0.6530265212059021, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 98.8339168978855, + "batch_question_ids_sha256": "8cebf4bd95bbe7d45c9169b299189db006a19e04aefa7bbb481fb6cbb5913227" + }, + { + "step": 126, + "phase": "full", + "loss": 0.7745178639888763, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 99.53647938836366, + "batch_question_ids_sha256": "01412ab449821f43118bb259e196bde1c079c22d33be200c8293ad3600a9f471" + }, + { + "step": 127, + "phase": "full", + "loss": 0.6720215082168579, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 100.24351753760129, + "batch_question_ids_sha256": "f358c70ad826fec50b070488b4a5e8cf15e1adacca80186e8b87ee614396e3b0" + }, + { + "step": 128, + "phase": "full", + "loss": 0.6795719712972641, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 100.94683667551726, + "batch_question_ids_sha256": "5f33904e2bd6ddd9725d283f6d401a2137795567327770ea0a4b0f9529d9d83c" + }, + { + "step": 129, + "phase": "full", + "loss": 0.6806273907423019, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 101.65407109819353, + "batch_question_ids_sha256": "f3b644027d2f16553adf3b5e697e434d7e25093ac988e6f860c07c4b98f42333" + }, + { + "step": 130, + "phase": "full", + "loss": 0.733887642621994, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 102.36723616439849, + "batch_question_ids_sha256": "e355a33e4a7d6f1eebbea569ec067a02b85b83584317c96c0a41374e61e69479" + }, + { + "step": 131, + "phase": "full", + "loss": 0.6697006225585938, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 103.13169718533754, + "batch_question_ids_sha256": "9bf6d8992b3affc7cd3e4018fe641b4afcec762703ed8aae75b3fa5034132e5c" + }, + { + "step": 132, + "phase": "full", + "loss": 0.6697288453578949, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 103.83426981512457, + "batch_question_ids_sha256": "59e4c200cfbc1a2e6193a0bd626d2fd6251eb9a5addf3313d80d0c8229c60d8c" + }, + { + "step": 133, + "phase": "full", + "loss": 0.6281428635120392, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 104.54028215818107, + "batch_question_ids_sha256": "f4b5288c511e28b22882d879371ff06d701ab148620b6d869e9e7cc270ae2ffa" + }, + { + "step": 134, + "phase": "full", + "loss": 0.6243098452687263, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 105.24348337762058, + "batch_question_ids_sha256": "a7879986f186e6a3ecf5f96ed8995fb9a51b757a2c3be32fd7d35fbac305bf56" + }, + { + "step": 135, + "phase": "full", + "loss": 0.6557051539421082, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 105.95341565273702, + "batch_question_ids_sha256": "84c6022baa2f94b378077577bcca8bb86162d35ad1f7044350b278c2e422ca29" + }, + { + "step": 136, + "phase": "full", + "loss": 0.7944526821374893, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 106.65815098118037, + "batch_question_ids_sha256": "4ebc0b2dee40a54f97dea45b41798dc30c0f5c79965bc7808ac3dd86de3ea5a5" + }, + { + "step": 137, + "phase": "full", + "loss": 0.8478172421455383, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 107.36669033393264, + "batch_question_ids_sha256": "e15c97134b0a010f04c895522470b1a0e69012df786a54ce556c5aa02290d442" + }, + { + "step": 138, + "phase": "full", + "loss": 0.6588033437728882, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 108.07306809257716, + "batch_question_ids_sha256": "9053897120d0b7ebbdfad5ed7c991a6b988b4100ce873610882f91bd09fd5459" + }, + { + "step": 139, + "phase": "full", + "loss": 0.7283994629979134, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 108.78070947993547, + "batch_question_ids_sha256": "efee6c6600f2549036d1c8ca99c670f8ea54c4c2bdbc56516d7b1f11a362fffc" + }, + { + "step": 140, + "phase": "full", + "loss": 0.6958236545324326, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 109.48458141926676, + "batch_question_ids_sha256": "3c1a629b79c1fd90b8b19ebf21f665f6cf6d592de4ab10c697dfcc86b8532340" + }, + { + "step": 141, + "phase": "full", + "loss": 0.764183297753334, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 110.19124841317534, + "batch_question_ids_sha256": "0ee316951cd9d9a97fa481cc459fbc51c49bd73f76975392712c0dcd0e28f41d" + }, + { + "step": 142, + "phase": "full", + "loss": 0.6614540070295334, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 110.89516895730048, + "batch_question_ids_sha256": "a8a1f161cf3a04612f95244bbf079956405b372b293d20ae1aa6e2456de55616" + }, + { + "step": 143, + "phase": "full", + "loss": 0.6594480872154236, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 111.61089933849871, + "batch_question_ids_sha256": "9d45e44c7cd61a4d4e968ecafe537b4c8a9be78edfc4d9480712465a14291906" + }, + { + "step": 144, + "phase": "full", + "loss": 0.6705136299133301, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 112.37249011918902, + "batch_question_ids_sha256": "203125e9a3222c2c0379c28f501b266a670c1ae9ba85505822cfaa52b2edb93f" + }, + { + "step": 145, + "phase": "full", + "loss": 0.6748034507036209, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 113.08026526775211, + "batch_question_ids_sha256": "82bd7004933e75cc6d8897b7be9e830ecb87d0dd83551ef94e16babe30a1626d" + }, + { + "step": 146, + "phase": "full", + "loss": 0.6681219041347504, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 113.78349374979734, + "batch_question_ids_sha256": "608a72b245a30555b5be3ada5d8108714f3c91f90c0a7cffd107b3b084f82eb8" + }, + { + "step": 147, + "phase": "full", + "loss": 0.6792533099651337, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 114.49104763753712, + "batch_question_ids_sha256": "0cdcbbd63e8c8045dc28b482951c730db5f58c693f94462bb4746465f0440b97" + }, + { + "step": 148, + "phase": "full", + "loss": 0.6486925780773163, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 115.19401026889682, + "batch_question_ids_sha256": "8d9e439b5d2f449feeb51745b427a817594ab81dc359e40a05316885001d8bc0" + }, + { + "step": 149, + "phase": "full", + "loss": 0.6756344139575958, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 115.90279173478484, + "batch_question_ids_sha256": "4b943eeb601bded64b4a1453c10c3b0c36b8625c278e3c950d07ffe9f7fc1ff7" + }, + { + "step": 150, + "phase": "full", + "loss": 0.6134026497602463, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 116.60606201924384, + "batch_question_ids_sha256": "7ff7570ca4b5adb924848173d4defd963987c68576596cf45e0560b9e5eb38fb", + "dev": { + "questions": 192, + "objective": "gold_distribution", + "eligible_questions": 192, + "target_ce": 0.6790979170447179, + "target_kl": 0.6790979170447179, + "target_tv": 0.4891960062803631, + "by_target_kind": { + "deterministic_truth": { + "n": 192, + "ce": 0.6790979170447179, + "kl": 0.6790979170447179, + "tv": 0.4891960062803631 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 192, + "ce": 0.6790979170447179, + "kl": 0.6790979170447179, + "tv": 0.4891960062803631 + }, + "observed_outcome": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + { + "step": 151, + "phase": "full", + "loss": 0.6479852646589279, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 118.55705555528402, + "batch_question_ids_sha256": "a37cef2c4dec236e9407b98d61862b07896b52c704e85cef4b32aff18df2d00b" + }, + { + "step": 152, + "phase": "full", + "loss": 0.6288465261459351, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 119.26285979244858, + "batch_question_ids_sha256": "51dc3787b6fe767ee279870d42b25e65efc1be7d08f6a5266e1d58048fbed773" + }, + { + "step": 153, + "phase": "full", + "loss": 0.6410926282405853, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 119.97168881259859, + "batch_question_ids_sha256": "9346ec3ed99488f9982d1073fc29d6ae7317741f29d6e81dabbd1c428229c3b1" + }, + { + "step": 154, + "phase": "full", + "loss": 0.5813655108213425, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 120.6752430703491, + "batch_question_ids_sha256": "72adc1de150d7168af2216acb891bbd1f1e608ed1b071b3b5a8aa2f77b015d26" + }, + { + "step": 155, + "phase": "full", + "loss": 0.6355779021978378, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 121.3853417718783, + "batch_question_ids_sha256": "0363b5b270df508617350e8a2e6cfd221f1f9ea3b4522912207ae91dc66f7ecd" + }, + { + "step": 156, + "phase": "full", + "loss": 0.8860462158918381, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 122.09786409232765, + "batch_question_ids_sha256": "1e6785a661a721048bfab1bca97d5d125e99e2673623b588c3a5087335f4e31c" + }, + { + "step": 157, + "phase": "full", + "loss": 0.8420947194099426, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 122.86192176491022, + "batch_question_ids_sha256": "c658995ff43878c0131af31f7452918cb88b0546b915ce641d2b3c0ea5495b8a" + }, + { + "step": 158, + "phase": "full", + "loss": 0.8676016628742218, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 123.56715064682066, + "batch_question_ids_sha256": "881aaeb396b5e6d129f57eb8c0d618ad6c3e3f92ec94c98db83aee4d7fc564cc" + }, + { + "step": 159, + "phase": "full", + "loss": 0.6374461725354195, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 124.27640721201897, + "batch_question_ids_sha256": "eedad14cef3952d1db431c64329ad96a9a3c942260d7254ee507ab50e6018e26" + }, + { + "step": 160, + "phase": "full", + "loss": 0.7472737208008766, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 124.98104074038565, + "batch_question_ids_sha256": "e71b19d0f5864d5bdd3447e82321e1497f88dd5786f436476a8839ad695e726c" + }, + { + "step": 161, + "phase": "full", + "loss": 0.6023682132363319, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 125.69018721114844, + "batch_question_ids_sha256": "e72e4cc1014af1eb5f88fda577c4b35068ea3ddaccc0046d2df1cbd0523663ae" + }, + { + "step": 162, + "phase": "full", + "loss": 0.7771216258406639, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 126.39480255357921, + "batch_question_ids_sha256": "455536c7e34c2ab56a1fda247fe6a5af66ab1ed5524cf72793a22685d624bef2" + }, + { + "step": 163, + "phase": "full", + "loss": 0.6805876046419144, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 127.1033055158332, + "batch_question_ids_sha256": "51fbe8ad21831c4a4782b0c0c39930aa0c887bd695889c97ce48d9babe541112" + }, + { + "step": 164, + "phase": "full", + "loss": 0.6350379884243011, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 127.80822979193181, + "batch_question_ids_sha256": "00e1d181ea4959d3279060b872716639cbab71f4e055f309dd6b80cfd45cb6d3" + }, + { + "step": 165, + "phase": "full", + "loss": 0.7217883318662643, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 128.51988022681326, + "batch_question_ids_sha256": "4d349e0386fb20c285d27bc3f37d8872b6d17daf20776612982e1d878d7328da" + }, + { + "step": 166, + "phase": "full", + "loss": 0.6808428168296814, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 129.2229075562209, + "batch_question_ids_sha256": "970185e82c151c8bb7936b6530c7f62142752bbb5a9def513f404c0d872583ab" + }, + { + "step": 167, + "phase": "full", + "loss": 0.6799975037574768, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 129.93164676986635, + "batch_question_ids_sha256": "1e2e4fcbd41ccd0a56f492ba1a08cee9d15a6c373901c3096311ff85aed42b5f" + }, + { + "step": 168, + "phase": "full", + "loss": 0.6509194672107697, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 130.63518224097788, + "batch_question_ids_sha256": "29baff208dd9d8ee64916b31412d6b2a3a83fe5156d19709cb123466ef67b032" + }, + { + "step": 169, + "phase": "full", + "loss": 0.6296563521027565, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 131.35125130414963, + "batch_question_ids_sha256": "aa118e2a9e2a13d820504e21f51b75f5d4a9c828508d72b77f6b44ca2ff5b455" + }, + { + "step": 170, + "phase": "full", + "loss": 0.6982598379254341, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 132.10903760790825, + "batch_question_ids_sha256": "3b3dece30df0e48a2c1600205dedaeef0fd03a49c7e1a4d337e8a86f48dc505d" + }, + { + "step": 171, + "phase": "full", + "loss": 0.5808834284543991, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 132.81755509786308, + "batch_question_ids_sha256": "86a1cc6990b94afe2827ca103bc01900af1e4f1e88e89831ce18abf1bd08051d" + }, + { + "step": 172, + "phase": "full", + "loss": 0.7441476136445999, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 133.52160083688796, + "batch_question_ids_sha256": "9e9ad0818bc099c7325fa7287a75da68d04a429cb7da04c876a7ce58b0fdb5e1" + }, + { + "step": 173, + "phase": "full", + "loss": 0.694560095667839, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 134.2304227631539, + "batch_question_ids_sha256": "9e06fd11bd1459acebbd48d4a4ce48e9357b6da34793875f674f53f8b70904a9" + }, + { + "step": 174, + "phase": "full", + "loss": 0.5658717751502991, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 134.93446967564523, + "batch_question_ids_sha256": "96879991783769359518d4caf27573932840e6e152064001b77338a6381ba097" + }, + { + "step": 175, + "phase": "full", + "loss": 0.7630509287118912, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 135.64201147854328, + "batch_question_ids_sha256": "c7f81c1973ce2d311a964472cd7c9ebaa7ea4d46bf0bec2ec70b4265e5835dd4" + }, + { + "step": 176, + "phase": "full", + "loss": 0.7058666348457336, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 136.34556037653238, + "batch_question_ids_sha256": "caba59efa93f32aaf55c668f3e2a9697d678ca82c5594afdc6069bbd7732f993" + }, + { + "step": 177, + "phase": "full", + "loss": 0.666761115193367, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 137.05348644126207, + "batch_question_ids_sha256": "4e85b5071b11eda85fbef981e7c984e0866d3ecd9fed9e6edd228b0427a426fd" + }, + { + "step": 178, + "phase": "full", + "loss": 0.6882964819669724, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 137.75807391945273, + "batch_question_ids_sha256": "2d1feea677f65f20366d7712d67f961d882b484b1135080d50640634f34695fc" + }, + { + "step": 179, + "phase": "full", + "loss": 0.6656223386526108, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 138.4652737043798, + "batch_question_ids_sha256": "6366245ec56686d7d8b8967bdb9612fe582ebc31513e18142bad06edeb13b143" + }, + { + "step": 180, + "phase": "full", + "loss": 0.6959905177354813, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 139.17376198712736, + "batch_question_ids_sha256": "09293e203503f77ea4562c92f0708e1f945a6ea9679d126d6a51af777ad35aca" + }, + { + "step": 181, + "phase": "full", + "loss": 0.6623962819576263, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 139.88249521050602, + "batch_question_ids_sha256": "0110c18cf76c6d3875bb5a574b727a4b02e0589eb64f9d49217c82e3c9ee0439" + }, + { + "step": 182, + "phase": "full", + "loss": 0.6377524137496948, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 140.59350656624883, + "batch_question_ids_sha256": "ee67052d7a9198f2dee5e73ac30fe190be6c3298f2c6e56d893ea1e48afd76ed" + }, + { + "step": 183, + "phase": "full", + "loss": 0.644087627530098, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 141.3565265405923, + "batch_question_ids_sha256": "f28d72dea28a603cfc4445a466c145154cf5be6cd3d1aff684666913973cf5e1" + }, + { + "step": 184, + "phase": "full", + "loss": 0.7233986705541611, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 142.0599279506132, + "batch_question_ids_sha256": "f30954cb2299e5cac4ca757e2d11840e3478108b99fc08d8329779fdfa8427af" + }, + { + "step": 185, + "phase": "full", + "loss": 0.6750015616416931, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 142.76889378763735, + "batch_question_ids_sha256": "98fa36fb81d9417f9a4209a509ba7a036b54cc1a993c5d257be4643ff53be127" + }, + { + "step": 186, + "phase": "full", + "loss": 0.69915971159935, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 143.47215379029512, + "batch_question_ids_sha256": "1bc982ff6cb294072bcd4ad98ccea4df10a48295b73243f4cfdee3b19fc11c4c" + }, + { + "step": 187, + "phase": "full", + "loss": 0.6474266946315765, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 144.1807238738984, + "batch_question_ids_sha256": "1eaa45386736be2d480969143abe0b2360fc1f59b701581906aa27f01efab036" + }, + { + "step": 188, + "phase": "full", + "loss": 0.7007929980754852, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 144.88448021747172, + "batch_question_ids_sha256": "a65d0059d031355321086ba85ad7ea89de275aaf15407d0ed0c859cb70a5795e" + }, + { + "step": 189, + "phase": "full", + "loss": 0.7222506254911423, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 145.59127262607217, + "batch_question_ids_sha256": "37a1b2ce577f9a54c8f56ab1d4d3946b8ad4800c7f824fb6e80b576a6ef3c3bd" + }, + { + "step": 190, + "phase": "full", + "loss": 0.6460579633712769, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 146.29686843510717, + "batch_question_ids_sha256": "9735ff0335ba3b5e86a0912ef078b934cef329c3a33db95e5be5afdf0e629d66" + }, + { + "step": 191, + "phase": "full", + "loss": 0.6607587188482285, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 147.0035694455728, + "batch_question_ids_sha256": "72c1377f10ffd6b6d9d2fe9c5b650be6e55acc3cef3276e2a5f2067246ca5ae0" + }, + { + "step": 192, + "phase": "full", + "loss": 0.5912963002920151, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 147.70575940608978, + "batch_question_ids_sha256": "ffb9020070f906dcad076864b7ccc310d2144ddc8a5b3e5fba2925fac69e16a4" + }, + { + "step": 193, + "phase": "full", + "loss": 0.69478739798069, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 148.41139391157776, + "batch_question_ids_sha256": "73a53ad0bb183aa2546250f1ffa85ae9d226eb56528823e9132eb14ff3c600fd" + }, + { + "step": 194, + "phase": "full", + "loss": 0.6415392011404037, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 149.11875948030502, + "batch_question_ids_sha256": "f32d1fa75c483cf6f5f9693baeb5f3716456eeaa1addf9f846d845cfd66ee242" + }, + { + "step": 195, + "phase": "full", + "loss": 0.6452083587646484, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 149.83714293688536, + "batch_question_ids_sha256": "bb249815e47bb1205c76ef35e124fb0ef9387070478ea28b05710649cca84243" + }, + { + "step": 196, + "phase": "full", + "loss": 0.6388794034719467, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 150.59404563345015, + "batch_question_ids_sha256": "67142093429c4cd962636d98cd7642af8d58a77451122e61b585f1b7272a0621" + }, + { + "step": 197, + "phase": "full", + "loss": 0.6328594237565994, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 151.30085357185453, + "batch_question_ids_sha256": "9f36f11d4f1e26318324388cb7290ae09fcab285da84c091fb216b496f99b98d" + }, + { + "step": 198, + "phase": "full", + "loss": 0.5657988637685776, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 152.00254012923688, + "batch_question_ids_sha256": "b9e0db1f2b8cc159af2ac50f2872aee7cc9438f5ad1ae81812378c519ec2f8aa" + }, + { + "step": 199, + "phase": "full", + "loss": 0.6964341253042221, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 152.70927072688937, + "batch_question_ids_sha256": "c934a0cba8a7a24570d888a6e63dacb0f9040ffaa7487808b8b4c49873f1eeb0" + }, + { + "step": 200, + "phase": "full", + "loss": 0.4618035778403282, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 153.41434872429818, + "batch_question_ids_sha256": "3e71aaf477d3ef20e5fce56fb2041df50d7cb4db5d5cd79feeb025c595192ada", + "dev": { + "questions": 192, + "objective": "gold_distribution", + "eligible_questions": 192, + "target_ce": 0.6276765607733715, + "target_kl": 0.6276765607733715, + "target_tv": 0.43645941461286886, + "by_target_kind": { + "deterministic_truth": { + "n": 192, + "ce": 0.6276765607733715, + "kl": 0.6276765607733715, + "tv": 0.43645941461286886 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 192, + "ce": 0.6276765607733715, + "kl": 0.6276765607733715, + "tv": 0.43645941461286886 + }, + "observed_outcome": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + { + "step": 201, + "phase": "full", + "loss": 0.7193554937839508, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 158.73019159585238, + "batch_question_ids_sha256": "9e85c4694e55f8e97437450924f30e6723f0c5fea57c58659b4f0c30ff0c7272" + }, + { + "step": 202, + "phase": "full", + "loss": 0.6497115790843964, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 159.4372734799981, + "batch_question_ids_sha256": "98656f90bf6e513850f5dc223fccf1dc677b069ee692e5c80a40515433dbcd60" + }, + { + "step": 203, + "phase": "full", + "loss": 0.5829432681202888, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 160.14536030963063, + "batch_question_ids_sha256": "29e733dac830059aa7cfbff3165340ca37f40236647455bb5c7df30b8701fb95" + }, + { + "step": 204, + "phase": "full", + "loss": 0.7340008914470673, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 160.85061140544713, + "batch_question_ids_sha256": "b4726f8928bb516f874ec0945c05313a170fb8f3f63f41fdde434d490e8a2a46" + }, + { + "step": 205, + "phase": "full", + "loss": 0.6012581139802933, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 161.559972865507, + "batch_question_ids_sha256": "32fddef8997617e5a2396b3120f53d4025d0ed8070a6efd3617465a8563f8bfc" + }, + { + "step": 206, + "phase": "full", + "loss": 0.5873189866542816, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 162.26373360678554, + "batch_question_ids_sha256": "155b5d5d5c4a46a05d86699f124433a742e6f515a45c358606c8208b68962c59" + }, + { + "step": 207, + "phase": "full", + "loss": 0.4397193491458893, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 162.97328280564398, + "batch_question_ids_sha256": "ce329c54c9edc137d2954e9d14ebcd58676a004bd26990f45a934c922a9d3997" + }, + { + "step": 208, + "phase": "full", + "loss": 0.5868460685014725, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 163.68661765288562, + "batch_question_ids_sha256": "44ccf1809c284b767d98a48c56705cebdfd43a127eab77886521fdd726b3239c" + }, + { + "step": 209, + "phase": "full", + "loss": 0.5234366804361343, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 164.45213057100773, + "batch_question_ids_sha256": "92ff6ae7092a2167fd51602bcf019eae9312436fc048be37a452f282d877c944" + }, + { + "step": 210, + "phase": "full", + "loss": 0.6880384087562561, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 165.15703650098294, + "batch_question_ids_sha256": "61446b75cb8ee3b664be4bc30865ad911574d44df45af89b8842029ea2a3bdf6" + }, + { + "step": 211, + "phase": "full", + "loss": 0.5493590906262398, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 165.86391623038799, + "batch_question_ids_sha256": "2bb8ad1136020ce7724c1e09fbcf4551a20ac54bf6da845ed6789a82e3200c2e" + }, + { + "step": 212, + "phase": "full", + "loss": 0.5554358810186386, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 166.56692613475025, + "batch_question_ids_sha256": "228dc4d5ed19561b4b3ffeae7c9df6e68eef2c7d3ae673a19595179fcdb0d2d2" + }, + { + "step": 213, + "phase": "full", + "loss": 0.37986456602811813, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 167.2739802254364, + "batch_question_ids_sha256": "9b39e8629c63bbe59e427459850a6a1a6bf7c4721e07bdea8c8daf187d16b11a" + }, + { + "step": 214, + "phase": "full", + "loss": 0.5644498988986015, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 167.97674071136862, + "batch_question_ids_sha256": "63f5c16a385b98e752e3a3eb61f7769b659ec1072079573b6a067c1ad8df9c64" + }, + { + "step": 215, + "phase": "full", + "loss": 0.5414157435297966, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 168.68344496749341, + "batch_question_ids_sha256": "5e71ab0ad8e1ae3d8e8c1751397b1335e23c70ef567d57c9665756d3d7f945a3" + }, + { + "step": 216, + "phase": "full", + "loss": 0.5384631305932999, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 169.38694399781525, + "batch_question_ids_sha256": "db22c42da509b8693c49512ebae718f5aef71ebe56817fa08c76b93bcebaf096" + }, + { + "step": 217, + "phase": "full", + "loss": 0.5437204763293266, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 170.0997542105615, + "batch_question_ids_sha256": "17802e36d2f622fd4c8885f4f4ce458475059f0a6308fbd8d70e335ef6ae6557" + }, + { + "step": 218, + "phase": "full", + "loss": 0.4782334491610527, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 170.8092891704291, + "batch_question_ids_sha256": "b8ddf9f4624a636ccd25e0d8d2c69d6c365e2a6859d41557488dc106f55b383c" + }, + { + "step": 219, + "phase": "full", + "loss": 0.4779422804713249, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 171.5236063785851, + "batch_question_ids_sha256": "6c2d2a20e601905e06172aa7fc56b5b0a9e884d2ff1cc75bfd324011e0d189dd" + }, + { + "step": 220, + "phase": "full", + "loss": 0.6652021333575249, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 172.2323320666328, + "batch_question_ids_sha256": "bf666f4387d3147cacf71021b0b6b2a5d5489a87b9a3cb2a7eac997a93fc9e98" + }, + { + "step": 221, + "phase": "full", + "loss": 0.40736595541238785, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 172.95417359285057, + "batch_question_ids_sha256": "71467a773e0feb4bda36d68952851755e756fc90961613854dcbeb6eb85777ec" + }, + { + "step": 222, + "phase": "full", + "loss": 0.2436215654015541, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 173.72030964959413, + "batch_question_ids_sha256": "54b6b00d1b157ece173add1c111bd04fee9eaa69fa4d0db455c32a7576334e74" + }, + { + "step": 223, + "phase": "full", + "loss": 0.3627225011587143, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 174.4337683096528, + "batch_question_ids_sha256": "5d34dc12370ef159339cfdf6f3bb0e9b50c460cb6689d9c2bb4a9660b7fba3ae" + }, + { + "step": 224, + "phase": "full", + "loss": 0.48189232498407364, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 175.1425352273509, + "batch_question_ids_sha256": "4df2bc9301d5389023f44a42c68215c745ead96869c6ce7d58219c54044ccb6a" + }, + { + "step": 225, + "phase": "full", + "loss": 0.3397899866104126, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 175.85545202810317, + "batch_question_ids_sha256": "f87d52165238d94c57f703976bb4e0ab754ea3b1519a810107a6cae356a0a488" + }, + { + "step": 226, + "phase": "full", + "loss": 0.37861377373337746, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 176.56356605980545, + "batch_question_ids_sha256": "7e00a9e1a3682a2bf3906f0df77600302f47d2b5f561f599b3b43a0c708a44fc" + }, + { + "step": 227, + "phase": "full", + "loss": 0.41144601441919804, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 177.2766804676503, + "batch_question_ids_sha256": "4a80e133b8d7a6d716cada1680ccd4dee163b2c18f0bf81904327233b1e8393e" + }, + { + "step": 228, + "phase": "full", + "loss": 0.373646242544055, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 177.98688609804958, + "batch_question_ids_sha256": "0b47464a00ced252390151a4cbc6fa85c3487ecca6ca2bbd130871dba5f5284a" + }, + { + "step": 229, + "phase": "full", + "loss": 0.48118457943201065, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 178.69982174504548, + "batch_question_ids_sha256": "09e81512f223e18f77e7e05c9dca3cb31f58c255fa4682ef3c4880a4bfb3e8f2" + }, + { + "step": 230, + "phase": "full", + "loss": 0.5880494406446815, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 179.40780649427325, + "batch_question_ids_sha256": "15545dc26bc4cde5b3005b6a7569d930dc727153644c166fa5481e635ce6ff88" + }, + { + "step": 231, + "phase": "full", + "loss": 0.351172374561429, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 180.12308822851628, + "batch_question_ids_sha256": "a338c2fb15e3b415edc457ece9a9faedab71f9169958b6593bbf9c802b2e6212" + }, + { + "step": 232, + "phase": "full", + "loss": 0.34208861738443375, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 180.8315440285951, + "batch_question_ids_sha256": "0841c4d53018f8164aaa1b78265c375de0981884ff4e569b4e4955b3510cb888" + }, + { + "step": 233, + "phase": "full", + "loss": 0.4977984121069312, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 181.54567025322467, + "batch_question_ids_sha256": "902e87f88199c0b6b025f92375bdb12d31cb3703cadad3dbe5fbd60ac9c6ad9f" + }, + { + "step": 234, + "phase": "full", + "loss": 0.21341162733733654, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 182.26411435939372, + "batch_question_ids_sha256": "d55f94bac50ccb152038b892097a9cd6820ee599734049d7facdddbf2d783e06" + }, + { + "step": 235, + "phase": "full", + "loss": 0.7171636074781418, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 183.03223790973425, + "batch_question_ids_sha256": "85a05bf651d9de9c2ed2a94d81fdfb73575c410ec35b168ec0881dd531e27a37" + }, + { + "step": 236, + "phase": "full", + "loss": 0.302105862647295, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 183.74024910107255, + "batch_question_ids_sha256": "6a9e0cdae4adda178099417f741733f56f2228eb8729905db9ff25cd9a84a8ef" + }, + { + "step": 237, + "phase": "full", + "loss": 0.21247326862066984, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 184.45327905286103, + "batch_question_ids_sha256": "e2222a864e75da183a27b0ee12c3893802d267113c8a35eb1264a7a46d962477" + }, + { + "step": 238, + "phase": "full", + "loss": 0.43014370277523994, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 185.1611839858815, + "batch_question_ids_sha256": "b53c2b1438b1702a807f93735cec535589011f0cf5f1cf562543fd0224cd3dc6" + }, + { + "step": 239, + "phase": "full", + "loss": 0.3218426760286093, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 185.87637827917933, + "batch_question_ids_sha256": "b5a1fb7b36a46761d4c39833591e99cc2876dc2bafc5505946964d3d5d14a2bb" + }, + { + "step": 240, + "phase": "full", + "loss": 0.5234263241291046, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 186.59336013626307, + "batch_question_ids_sha256": "0b3b43e53c9ef257aef9a0c9fc90b9dd4fb51e5b2fc9e1da40efc1126fd6b388" + }, + { + "step": 241, + "phase": "full", + "loss": 0.3241303414106369, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 187.30394040979445, + "batch_question_ids_sha256": "2360c9f9f4b22ebffc22082f6d3585815f72593f14aa0d1f5ae7699ff244626d" + }, + { + "step": 242, + "phase": "full", + "loss": 0.31431591510772705, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 188.01890878099948, + "batch_question_ids_sha256": "f28e247aaaeb7f6719351215472142dcedc59b6c8e4f96eae53a5b37dd31d586" + }, + { + "step": 243, + "phase": "full", + "loss": 0.33552829921245575, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 188.72786663193256, + "batch_question_ids_sha256": "266ce6235051352859b662b67ec031617d95798f84b72bbe1d842c1459a0beca" + }, + { + "step": 244, + "phase": "full", + "loss": 0.5181017033755779, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 189.44423062261194, + "batch_question_ids_sha256": "6f06108da38e24f993a189c3b02dcef67b91ad0b1bcf326743c5af935b0fc555" + }, + { + "step": 245, + "phase": "full", + "loss": 0.32815761491656303, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 190.15867064055055, + "batch_question_ids_sha256": "21f362fc032c9bed9df4bbf1a5d707e61588456c6cd795a3dc6b6c8079ad42d2" + }, + { + "step": 246, + "phase": "full", + "loss": 0.45874930918216705, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 190.8708414118737, + "batch_question_ids_sha256": "c904e5b7d2272756841cf377499bd81fe67b90100a618a7ca2dfeb41f1e28729" + }, + { + "step": 247, + "phase": "full", + "loss": 0.5854355692863464, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 191.5900476584211, + "batch_question_ids_sha256": "2eeba040f81866175ebf341837e30b96ca34c3615510ae938f9bf0b3c27b7bb6" + }, + { + "step": 248, + "phase": "full", + "loss": 0.7957623600959778, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 192.3616753295064, + "batch_question_ids_sha256": "dea653a95346ff74d8d19168dbc7063217b960aef7a629826c1257e45a1ddc85" + }, + { + "step": 249, + "phase": "full", + "loss": 0.46265869215130806, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 193.07118207402527, + "batch_question_ids_sha256": "0a6b42c33ca52755a4ec36ff16087a2d818821d700012167ad42e5da45b19b31" + }, + { + "step": 250, + "phase": "full", + "loss": 0.3855026252567768, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 193.78607384487987, + "batch_question_ids_sha256": "f5bdf5cf9212c0e520f9271cc9f3ff63d8f713b191a0ffc8b74c12a3d787de91", + "dev": { + "questions": 192, + "objective": "gold_distribution", + "eligible_questions": 192, + "target_ce": 0.5133817775634377, + "target_kl": 0.5133817775634377, + "target_tv": 0.3455996543839961, + "by_target_kind": { + "deterministic_truth": { + "n": 192, + "ce": 0.5133817775634377, + "kl": 0.5133817775634377, + "tv": 0.3455996543839961 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 192, + "ce": 0.5133817775634377, + "kl": 0.5133817775634377, + "tv": 0.3455996543839961 + }, + "observed_outcome": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + }, + { + "step": 251, + "phase": "full", + "loss": 0.6043301150202751, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 198.8792834635824, + "batch_question_ids_sha256": "5015d8b585bb8cf06e50c631566ad27f8a4eeded320db68560f7f0b151c278b6" + }, + { + "step": 252, + "phase": "full", + "loss": 0.7012896165251732, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 199.58792897406965, + "batch_question_ids_sha256": "318e877c3445d72ffe7991326de572f45929c825f27d67d32101b3c23027ff9b" + }, + { + "step": 253, + "phase": "full", + "loss": 0.7492517530918121, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 200.30492838658392, + "batch_question_ids_sha256": "326d4567b84254c1fc97da3b785e286e66ba070150a5ad4508c88125e62db961" + }, + { + "step": 254, + "phase": "full", + "loss": 0.4353172145783901, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 201.01396651379764, + "batch_question_ids_sha256": "6f5d215bd2f2701800663f40ddc26b050557f3e020ece3e0eb6376a8dbf60579" + }, + { + "step": 255, + "phase": "full", + "loss": 0.5577676184475422, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 201.72927809599787, + "batch_question_ids_sha256": "0be60a4c771cc96f52ccfa923149f1e81fcfb2736b1c39883cd4a6c99c36bca2" + }, + { + "step": 256, + "phase": "full", + "loss": 0.2988160718232393, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 202.43849961273372, + "batch_question_ids_sha256": "d710876cd61bdac07a68e4de0010ab50f7e511e41c0a74c462a4817c0c727f01" + }, + { + "step": 257, + "phase": "full", + "loss": 0.4778193086385727, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 203.15040820278227, + "batch_question_ids_sha256": "87e284110bb43b94c6a633759a8ee213dbd6e7e5d84cc41b5986ab153cac1b00" + }, + { + "step": 258, + "phase": "full", + "loss": 0.2608841732144356, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 203.85997896362096, + "batch_question_ids_sha256": "0ad73820086e527ffe549948eab1ad1215e4fddbb2889ce02b155a7297cb208f" + }, + { + "step": 259, + "phase": "full", + "loss": 0.36815424263477325, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 204.57307406235486, + "batch_question_ids_sha256": "6a888596ca5bfca5bd1106322e1c4ff2b5e5221a63becf1e4d8f71feeb97d4ef" + }, + { + "step": 260, + "phase": "full", + "loss": 0.25392908975481987, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 205.29338589962572, + "batch_question_ids_sha256": "5a81b8c0d2156d0bad48ca91ad4ddb43a17d00ca016b1c4efc8c349ede7dc826" + }, + { + "step": 261, + "phase": "full", + "loss": 0.39711093343794346, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 206.06153078190982, + "batch_question_ids_sha256": "8ac66f1706c618dc7449355d30bc72937ccef61f606e2eb82b7fe45a67e816d5" + }, + { + "step": 262, + "phase": "full", + "loss": 0.5015534572303295, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 206.77086908742785, + "batch_question_ids_sha256": "219f93e668a33e596af605826dbd3b88fa50975a5ab5aeda598212e9c15f5f39" + }, + { + "step": 263, + "phase": "full", + "loss": 0.3130811285227537, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 207.4844183344394, + "batch_question_ids_sha256": "6dcb389d177404607743e497fdc55a37fa2fc564887294a067cb46b8f373d3bc" + }, + { + "step": 264, + "phase": "full", + "loss": 0.8013373166322708, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 208.19425421208143, + "batch_question_ids_sha256": "f5e7ac0f0b45c158e9167197ba652d3c385aebc47cc46e7bf717ab3cabbefb28" + }, + { + "step": 265, + "phase": "full", + "loss": 0.25049281027168036, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 208.90834939759225, + "batch_question_ids_sha256": "2b324b68be3304ba59787de5b92021ed6a31ea367ff68b148caa4e06f3782904" + }, + { + "step": 266, + "phase": "full", + "loss": 0.17060182429850101, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 209.61890569329262, + "batch_question_ids_sha256": "cdce24996769ea8a6c8a603540f063b3fa99074754103df020a30456935d68c3" + }, + { + "step": 267, + "phase": "full", + "loss": 0.4261407256126404, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 210.33242365904152, + "batch_question_ids_sha256": "1c800df45c5022efb522674429ae9a8c02987b3fe389ae92edae1fabac919fcb" + }, + { + "step": 268, + "phase": "full", + "loss": 0.5302892606705427, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 211.05312117934227, + "batch_question_ids_sha256": "e270dddaf536ef6fbaa343e844a40987b30877f34f22dfe4667f20e90acf989e" + }, + { + "step": 269, + "phase": "full", + "loss": 0.40489593520760536, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 211.76157358288765, + "batch_question_ids_sha256": "ce65ccd0c9aecc5a23c3532574b0377ef774ecd231c9ed848d3ba650dc2d4dad" + }, + { + "step": 270, + "phase": "full", + "loss": 0.4067187234759331, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 212.4743232568726, + "batch_question_ids_sha256": "8243355e6b5f05fe679697b90e0cfd244a99e219affd877e71ab03d79ab9754b" + }, + { + "step": 271, + "phase": "full", + "loss": 0.4890149012207985, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 213.18467714823782, + "batch_question_ids_sha256": "682a037fc8c769b73de8d5082432b3bf85d0dc3145560aef0a25f92abbd1ba50" + }, + { + "step": 272, + "phase": "full", + "loss": 0.1917540356516838, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 213.89931836910546, + "batch_question_ids_sha256": "82471699d68d13747293bbb99e265f728c25d362653a061d2e6980063da8ce0e" + }, + { + "step": 273, + "phase": "full", + "loss": 0.5509352870285511, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 214.61755146179348, + "batch_question_ids_sha256": "adc0c521e9fa3a39f7fb639efa0368b1c0e19807b12e7273001c148e0fdd120f" + }, + { + "step": 274, + "phase": "full", + "loss": 0.3338756188750267, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 215.38510585110635, + "batch_question_ids_sha256": "d420226ef4a8c79a690185838f8e092f75ddae423ba87fe84c59b70ceae642e8" + }, + { + "step": 275, + "phase": "full", + "loss": 0.3800750896334648, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 216.0923615153879, + "batch_question_ids_sha256": "9379b7cdd9123895038ea2a1b8d01673d33377b8f3f677e595134d033e5136e0" + }, + { + "step": 276, + "phase": "full", + "loss": 0.4598808139562607, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 216.8066939432174, + "batch_question_ids_sha256": "47678dd1d63d83e3a1794473f88a8b8338420772d5a108309f604d849cfc041e" + }, + { + "step": 277, + "phase": "full", + "loss": 0.3033773899078369, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 217.51723690330982, + "batch_question_ids_sha256": "3d44da5bf6eeaa6613468b99be70a25e1ccf08cb3b565708268989ff06e842f7" + }, + { + "step": 278, + "phase": "full", + "loss": 0.4007577896118164, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 218.23152319155633, + "batch_question_ids_sha256": "299f20a64f0531df00500c2a4e5eee923870e003a25de430494799e7882f4ef0" + }, + { + "step": 279, + "phase": "full", + "loss": 0.3342380002140999, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 218.94039700832218, + "batch_question_ids_sha256": "57c036baa3ee5df0caa158f373e0ea0722fdd3332237ac6828b13ea36a5d2ecf" + }, + { + "step": 280, + "phase": "full", + "loss": 0.337104894220829, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 219.6539764655754, + "batch_question_ids_sha256": "98bca7b692082145b8e4609d17fe3de225e15d125e1ea92d0ecae5f50bd34ee1" + }, + { + "step": 281, + "phase": "full", + "loss": 0.22853449173271656, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 220.3643310572952, + "batch_question_ids_sha256": "87c0078e6b2f02c7f957d0b63bbf840109765ed52c84960507e13bcab8c6157e" + }, + { + "step": 282, + "phase": "full", + "loss": 0.5159055069088936, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 221.07112131081522, + "batch_question_ids_sha256": "41a438ad7a7862108850774e4865d423e6f2778d66e9fa8221236211a71efe24" + }, + { + "step": 283, + "phase": "full", + "loss": 0.8988195657730103, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 221.77962809149176, + "batch_question_ids_sha256": "7ce2c829eef2ab4e26a79243d0bc8b2df073395c59e2524f9b83926021e4fab7" + }, + { + "step": 284, + "phase": "full", + "loss": 0.5292624160647392, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 222.4944567391649, + "batch_question_ids_sha256": "60d21ffdfc6a350374d5d2cf8aabb2bc8786ab93566517bb4831e47258a89e3e" + }, + { + "step": 285, + "phase": "full", + "loss": 0.2911728471517563, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 223.20228900015354, + "batch_question_ids_sha256": "a450ea269b3839972bd87373bbf6e2b65a3406834b4212377fe54397ca9b685c" + }, + { + "step": 286, + "phase": "full", + "loss": 0.9477544538676739, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 223.92744197975844, + "batch_question_ids_sha256": "0e6b6d1f05ab0a8cb6e36df37c25e449161ea5579265da98e8563d390f0257da" + }, + { + "step": 287, + "phase": "full", + "loss": 0.6072863787412643, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 224.69084625598043, + "batch_question_ids_sha256": "46fe51009bed98c1dce34ce9c3da24a83407b35af499e27e05a1e59c4a5ebf8e" + }, + { + "step": 288, + "phase": "full", + "loss": 0.314976304769516, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 225.40723146684468, + "batch_question_ids_sha256": "79ea50f61028a5d6955062e4ae0397ce61a9fd5dc8620cbda72ead8f28db6fa6" + }, + { + "step": 289, + "phase": "full", + "loss": 0.5241311118006706, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 226.11960257124156, + "batch_question_ids_sha256": "f1a7b181a19b0b4c218cc59a21bcc8818d3a3d10993de8c4dbd29b98e0d23f25" + }, + { + "step": 290, + "phase": "full", + "loss": 0.32841183990240097, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 226.83277582284063, + "batch_question_ids_sha256": "19c53fd533d3098690442155ef27a6bf2e7bc0c949a378b029301c5514eb07b6" + }, + { + "step": 291, + "phase": "full", + "loss": 0.7611794248223305, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 227.5416121473536, + "batch_question_ids_sha256": "6378a12d0d0baa58cce25fa07bbb11cb0d3acd6221f59dab6c6a64e452a7a31f" + }, + { + "step": 292, + "phase": "full", + "loss": 0.7156968712806702, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 228.25686148833483, + "batch_question_ids_sha256": "796cda091f0dc7c429af2a4668669c2bcf9dba1ca5ceee6ae2e72ad8e62ee76b" + }, + { + "step": 293, + "phase": "full", + "loss": 0.5687246583402157, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 228.96636441536248, + "batch_question_ids_sha256": "12b564664bc88fa037eb58b2c6c275c38d8a3b9c7c21c529351518578d7f812b" + }, + { + "step": 294, + "phase": "full", + "loss": 0.47587039694190025, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 229.68278923351318, + "batch_question_ids_sha256": "01db358b1c58a9ffab2e8a3f48bdb308be03029bf790dff03ab92be7ba7ef99a" + }, + { + "step": 295, + "phase": "full", + "loss": 0.5471991784870625, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 230.39351106341928, + "batch_question_ids_sha256": "1c1ac4b8abf469a972cfcf669d41e3ccbe576433ef786f0870624498b8a6e149" + }, + { + "step": 296, + "phase": "full", + "loss": 0.300723435357213, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 231.10899952147156, + "batch_question_ids_sha256": "d7776bf9456c83c12a480bc681a7575d385b257a8410c893570d75f48ef0e0f9" + }, + { + "step": 297, + "phase": "full", + "loss": 0.5512101650238037, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 231.81829000357538, + "batch_question_ids_sha256": "e7bbc683007eb5abce20d06fdcb109ce6e198abbfdfc403576d9f06ffcfde6ca" + }, + { + "step": 298, + "phase": "full", + "loss": 0.4343995898962021, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 232.5304178809747, + "batch_question_ids_sha256": "42b03d4d2bbbf3e2e336fb349e9555dae3f5cff56ce419d4520a7ced34a498b9" + }, + { + "step": 299, + "phase": "full", + "loss": 0.3628891035914421, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 233.2486968645826, + "batch_question_ids_sha256": "a128f7e7a4428c1e711df13b4350d3ca49bfd2526560538504755cbc6dad9acb" + }, + { + "step": 300, + "phase": "full", + "loss": 0.22461803443729877, + "questions": 16, + "microbatches": 4, + "elapsed_seconds": 234.0151388682425, + "batch_question_ids_sha256": "e0dbbe2cc8da87a9d012c0208a2afbdd9595a7397612d06164e5ebfa624c6b48", + "dev": { + "questions": 192, + "objective": "gold_distribution", + "eligible_questions": 192, + "target_ce": 0.48294000757494554, + "target_kl": 0.48294000757494554, + "target_tv": 0.31825824177139705, + "by_target_kind": { + "deterministic_truth": { + "n": 192, + "ce": 0.48294000757494554, + "kl": 0.48294000757494554, + "tv": 0.31825824177139705 + } + }, + "independent_targets": { + "teacher": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + }, + "gold_distribution": { + "n": 192, + "ce": 0.48294000757494554, + "kl": 0.48294000757494554, + "tv": 0.31825824177139705 + }, + "observed_outcome": { + "n": 0, + "ce": 0.0, + "kl": 0.0, + "tv": 0.0 + } + }, + "note": "question-mean distribution recovery; policy targets and reference argmax labels are not observed outcomes" + } + } +]