three-currency re-score with per-problem arrays and hull bootstrap

analysis/bayes_sc_2026-10-25/v8/bc8_gpu_s.arrays.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fa8cdd645c9394eee2b0e0745129094ca6ec6bb1725ebe2f58a661f7c2c4155
+size 39216

analysis/bayes_sc_2026-10-25/v8/bc8_gpu_s.json

@@ -0,0 +1,142 @@
+{
+ "cost_model": {
+  "currency": "gpu_s",
+  "unit": "GPU-seconds",
+  "per_equiv": 0.0006341154090044388,
+  "scenario": null,
+  "base": "qwen3-8b:nothink"
+ },
+ "ref": {
+  "tier": "think16k",
+  "acc": 0.9279941219691403,
+  "equiv": 5397.500408759384
+ },
+ "frontier": [
+  {
+   "R": 2000,
+   "acc": 0.7883908890521675,
+   "equiv": 1129.5711941564073
+  },
+  {
+   "R": 4000,
+   "acc": 0.7883908890521675,
+   "equiv": 1129.5711941564073
+  },
+  {
+   "R": 6000,
+   "acc": 0.7883908890521675,
+   "equiv": 1129.5711941564073
+  },
+  {
+   "R": 8000,
+   "acc": 0.7883908890521675,
+   "equiv": 1129.5711941564073
+  },
+  {
+   "R": 12000,
+   "acc": 0.7883908890521675,
+   "equiv": 1129.5711941564073
+  },
+  {
+   "R": 16000,
+   "acc": 0.837304502991498,
+   "equiv": 2834.403313721442
+  },
+  {
+   "R": 24000,
+   "acc": 0.8872152828802351,
+   "equiv": 4230.600976175624
+  },
+  {
+   "R": 32000,
+   "acc": 0.9125118085441378,
+   "equiv": 4900.948252343457
+  },
+  {
+   "R": 48000,
+   "acc": 0.9316678912564291,
+   "equiv": 5448.530278926885
+  }
+ ],
+ "matched": [
+  {
+   "seed": 0,
+   "acc": 0.9294636296840558,
+   "equiv": 5450.9879613091925,
+   "R": [
+    44381.251420165536,
+    49612.630326552375
+   ],
+   "delta_vs_agree2": [
+    0.0014695077149155045,
+    -0.005405689094153458,
+    0.008187257268814946
+   ],
+   "actions": {
+    "mid:think16k": 369
+   }
+  },
+  {
+   "seed": 1,
+   "acc": 0.9267345439277841,
+   "equiv": 5334.087495098382,
+   "R": [
+    50191.39018487453,
+    46282.92419459038
+   ],
+   "delta_vs_agree2": [
+    -0.0012595780413561446,
+    -0.008239739687204786,
+    0.0058268605017319135
+   ],
+   "actions": {
+    "mid:think16k": 335
+   }
+  },
+  {
+   "seed": 2,
+   "acc": 0.930985619817361,
+   "equiv": 5409.716893546653,
+   "R": [
+    50976.04938951666,
+    46873.10056327461
+   ],
+   "delta_vs_agree2": [
+    0.002991497848220847,
+    -0.003673769287288757,
+    0.009761729820510129
+   ],
+   "actions": {
+    "mid:think16k": 349
+   }
+  },
+  {
+   "seed": 3,
+   "acc": 0.9316154088380394,
+   "equiv": 5388.791856758997,
+   "R": [
+    43948.42366826647,
+    47852.21935253754
+   ],
+   "delta_vs_agree2": [
+    0.0036212868688989203,
+    -0.0027828802351212334,
+    0.010129106749239005
+   ],
+   "actions": {
+    "mid:think16k": 358
+   }
+  }
+ ],
+ "agree2_by_tier": {
+  "think16k": {
+   "acc": 0.9279941219691403,
+   "equiv": 5397.500408759384,
+   "delta_vs_ref": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  }
+ }
+}

analysis/bayes_sc_2026-10-25/v8/bc8_gpu_s.log

@@ -0,0 +1,17 @@
+2026-10-25 14:12:14,778 online_rl_train INFO: ThinkBooster math evaluator loaded
+reference agree2 -> think16k: acc 0.928 equiv 5398 | base 0.788 at 1130 | always think16k 0.948 at 8284
+cost model gpu_s: 1 equiv = 0.0006341 GPU-seconds
+agree2 -> think16k       acc 0.928 equiv   5398 (   +0) | vs reference +0.0 [+0.0, +0.0]
+R    2000: acc 0.788 equiv   1130
+R    4000: acc 0.788 equiv   1130
+R    6000: acc 0.788 equiv   1130
+R    8000: acc 0.788 equiv   1130
+R   12000: acc 0.788 equiv   1130
+R   16000: acc 0.837 equiv   2834
+R   24000: acc 0.887 equiv   4231
+R   32000: acc 0.913 equiv   4901
+R   48000: acc 0.932 equiv   5449
+parity: Bayes at agree2's cost acc 0.929 equiv 5451 (+53) | vs agree2 +0.1 [-0.5, +0.8] | R [44381, 49613] | actions {'mid:think16k': 369}
+split 1: Bayes at agree2's cost acc 0.927 equiv 5334 (-63) | vs agree2 -0.1 [-0.8, +0.6] | R [50191, 46283] | actions {'mid:think16k': 335}
+split 2: Bayes at agree2's cost acc 0.931 equiv 5410 (+12) | vs agree2 +0.3 [-0.4, +1.0] | R [50976, 46873] | actions {'mid:think16k': 349}
+split 3: Bayes at agree2's cost acc 0.932 equiv 5389 (-9) | vs agree2 +0.4 [-0.3, +1.0] | R [43948, 47852] | actions {'mid:think16k': 358}

analysis/bayes_sc_2026-10-25/v8/bc8_gpu_s.provenance.json

@@ -0,0 +1,98 @@
+{
+ "started_at": "2026-10-25T14:12:12+04:00",
+ "host": "ws006600",
+ "user": "vlad.smirnov",
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
+ "argv": [
+  "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_control.py",
+  "--pool-dir",
+  "workdir/selector/pool_mth_k8_chat",
+  "--uhead",
+  "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+  "--prm",
+  "workdir/selector/pool_mth_k8_chat/prm_selector.json",
+  "--prm-reads",
+  "workdir/selector/pool_mth_k8_chat/prm_reads.json",
+  "--step-tokens",
+  "<redacted>",
+  "--critics",
+  "agree",
+  "--task-prior",
+  "workdir/bayes/mth_task_prior.json",
+  "--tbins",
+  "3",
+  "--cost-model",
+  "gpu_s",
+  "--tier",
+  "think16k=workdir/layer_probe/think16k_tier.json@qwen3-8b:think",
+  "--out",
+  "workdir/bayes/v8/bc8_gpu_s.json"
+ ],
+ "git": {
+  "commit": "f935c73993aae0c1137fdec2866aef2005007331",
+  "branch": "HEAD",
+  "subject": "Per-problem arrays in the Bayesian controllers; paired bootstrap against the hull mixture in currency_table",
+  "dirty": false,
+  "dirty_files": []
+ },
+ "gpus": {
+  "visible": "",
+  "devices": [
+   {
+    "index": "0",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
+   },
+   {
+    "index": "1",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
+   }
+  ]
+ },
+ "packages": {
+  "python": "3.11.15",
+  "torch": "2.9.1",
+  "vllm": "0.15.1",
+  "transformers": "4.57.6",
+  "speculators": "0.4.0.dev49",
+  "peft": null,
+  "accelerate": "1.13.0",
+  "numpy": "2.2.6",
+  "huggingface_hub": "0.36.2",
+  "openai": "2.32.0",
+  "wandb": "0.25.1"
+ },
+ "env": {
+  "CUDA_VISIBLE_DEVICES": ""
+ },
+ "models": [],
+ "files": {
+  "pool": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
+   "bytes": 65136887,
+   "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
+  },
+  "uhead": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+   "bytes": 9369632,
+   "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
+  },
+  "prm": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/prm_selector.json",
+   "bytes": 9360982,
+   "sha256": "0be50791aa0d0af5d2f2d239f8fdf551919f716c2c84b295553de06be28a3fcd"
+  },
+  "tier:think16k": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
+   "bytes": 26666,
+   "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
+  }
+ },
+ "finished_at": "2026-10-25T14:12:31+04:00",
+ "wall_secs": 18.9
+}

analysis/bayes_sc_2026-10-25/v8/bc8_tokens.arrays.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7e8f9347986681a7d5bc15a13c5e1e8522cd048e20aea56da1b7f8c4da4c2af
+size 22883

analysis/bayes_sc_2026-10-25/v8/bc8_tokens.json

@@ -0,0 +1,142 @@
+{
+ "cost_model": {
+  "currency": "tokens",
+  "unit": "base tokens",
+  "per_equiv": 1.0,
+  "scenario": null,
+  "base": "qwen3-8b:nothink"
+ },
+ "ref": {
+  "tier": "think16k",
+  "acc": 0.9279941219691403,
+  "equiv": 4617.346383961372
+ },
+ "frontier": [
+  {
+   "R": 2000,
+   "acc": 0.7883908890521675,
+   "equiv": 1104.853783982366
+  },
+  {
+   "R": 4000,
+   "acc": 0.7883908890521675,
+   "equiv": 1104.853783982366
+  },
+  {
+   "R": 6000,
+   "acc": 0.7883908890521675,
+   "equiv": 1104.853783982366
+  },
+  {
+   "R": 8000,
+   "acc": 0.7883908890521675,
+   "equiv": 1104.853783982366
+  },
+  {
+   "R": 12000,
+   "acc": 0.8320037787341241,
+   "equiv": 2256.260165844442
+  },
+  {
+   "R": 16000,
+   "acc": 0.8729400650781987,
+   "equiv": 3147.0462894930197
+  },
+  {
+   "R": 24000,
+   "acc": 0.9096777579510863,
+   "equiv": 3944.4870998215597
+  },
+  {
+   "R": 32000,
+   "acc": 0.9225884328749868,
+   "equiv": 4257.002844547077
+  },
+  {
+   "R": 48000,
+   "acc": 0.9362338616563451,
+   "equiv": 4676.8411986984365
+  }
+ ],
+ "matched": [
+  {
+   "seed": 0,
+   "acc": 0.9361288968195656,
+   "equiv": 4655.3114411672095,
+   "R": [
+    48361.3395565794,
+    44732.65597091439
+   ],
+   "delta_vs_agree2": [
+    0.008134774850425108,
+    0.0022554319303033487,
+    0.014592736433294834
+   ],
+   "actions": {
+    "mid:think16k": 466
+   }
+  },
+  {
+   "seed": 1,
+   "acc": 0.9306182428886324,
+   "equiv": 4628.1536895140125,
+   "R": [
+    48097.040311758275,
+    50399.93104936121
+   ],
+   "delta_vs_agree2": [
+    0.002624120919491971,
+    -0.004043770336937124,
+    0.009711871523039768
+   ],
+   "actions": {
+    "mid:think16k": 451
+   }
+  },
+  {
+   "seed": 2,
+   "acc": 0.9324551275322766,
+   "equiv": 4598.262076204472,
+   "R": [
+    49825.98952869387,
+    47632.1174822072
+   ],
+   "delta_vs_agree2": [
+    0.004461005563136351,
+    -0.0022567439907630934,
+    0.011283719953815472
+   ],
+   "actions": {
+    "mid:think16k": 439
+   }
+  },
+  {
+   "seed": 3,
+   "acc": 0.9348168363598194,
+   "equiv": 4613.861110527973,
+   "R": [
+    45769.134996550725,
+    51481.074300937435
+   ],
+   "delta_vs_agree2": [
+    0.006822714390679124,
+    0.0006284769602183275,
+    0.013120604597459852
+   ],
+   "actions": {
+    "mid:think16k": 441
+   }
+  }
+ ],
+ "agree2_by_tier": {
+  "think16k": {
+   "acc": 0.9279941219691403,
+   "equiv": 4617.346383961372,
+   "delta_vs_ref": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  }
+ }
+}

analysis/bayes_sc_2026-10-25/v8/bc8_tokens.log

@@ -0,0 +1,17 @@
+2026-10-25 14:12:14,988 online_rl_train INFO: ThinkBooster math evaluator loaded
+reference agree2 -> think16k: acc 0.928 equiv 4617 | base 0.788 at 1105 | always think16k 0.948 at 6341
+cost model tokens: 1 equiv = 1 base tokens
+agree2 -> think16k       acc 0.928 equiv   4617 (   +0) | vs reference +0.0 [+0.0, +0.0]
+R    2000: acc 0.788 equiv   1105
+R    4000: acc 0.788 equiv   1105
+R    6000: acc 0.788 equiv   1105
+R    8000: acc 0.788 equiv   1105
+R   12000: acc 0.832 equiv   2256
+R   16000: acc 0.873 equiv   3147
+R   24000: acc 0.910 equiv   3944
+R   32000: acc 0.923 equiv   4257
+R   48000: acc 0.936 equiv   4677
+parity: Bayes at agree2's cost acc 0.936 equiv 4655 (+38) | vs agree2 +0.8 [+0.2, +1.5] | R [48361, 44733] | actions {'mid:think16k': 466}
+split 1: Bayes at agree2's cost acc 0.931 equiv 4628 (+11) | vs agree2 +0.3 [-0.4, +1.0] | R [48097, 50400] | actions {'mid:think16k': 451}
+split 2: Bayes at agree2's cost acc 0.932 equiv 4598 (-19) | vs agree2 +0.4 [-0.2, +1.1] | R [49826, 47632] | actions {'mid:think16k': 439}
+split 3: Bayes at agree2's cost acc 0.935 equiv 4614 (-3) | vs agree2 +0.7 [+0.1, +1.3] | R [45769, 51481] | actions {'mid:think16k': 441}

analysis/bayes_sc_2026-10-25/v8/bc8_tokens.provenance.json

@@ -0,0 +1,98 @@
+{
+ "started_at": "2026-10-25T14:12:12+04:00",
+ "host": "ws006600",
+ "user": "vlad.smirnov",
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
+ "argv": [
+  "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_control.py",
+  "--pool-dir",
+  "workdir/selector/pool_mth_k8_chat",
+  "--uhead",
+  "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+  "--prm",
+  "workdir/selector/pool_mth_k8_chat/prm_selector.json",
+  "--prm-reads",
+  "workdir/selector/pool_mth_k8_chat/prm_reads.json",
+  "--step-tokens",
+  "<redacted>",
+  "--critics",
+  "agree",
+  "--task-prior",
+  "workdir/bayes/mth_task_prior.json",
+  "--tbins",
+  "3",
+  "--cost-model",
+  "tokens",
+  "--tier",
+  "think16k=workdir/layer_probe/think16k_tier.json",
+  "--out",
+  "workdir/bayes/v8/bc8_tokens.json"
+ ],
+ "git": {
+  "commit": "f935c73993aae0c1137fdec2866aef2005007331",
+  "branch": "HEAD",
+  "subject": "Per-problem arrays in the Bayesian controllers; paired bootstrap against the hull mixture in currency_table",
+  "dirty": false,
+  "dirty_files": []
+ },
+ "gpus": {
+  "visible": "",
+  "devices": [
+   {
+    "index": "0",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
+   },
+   {
+    "index": "1",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
+   }
+  ]
+ },
+ "packages": {
+  "python": "3.11.15",
+  "torch": "2.9.1",
+  "vllm": "0.15.1",
+  "transformers": "4.57.6",
+  "speculators": "0.4.0.dev49",
+  "peft": null,
+  "accelerate": "1.13.0",
+  "numpy": "2.2.6",
+  "huggingface_hub": "0.36.2",
+  "openai": "2.32.0",
+  "wandb": "0.25.1"
+ },
+ "env": {
+  "CUDA_VISIBLE_DEVICES": ""
+ },
+ "models": [],
+ "files": {
+  "pool": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
+   "bytes": 65136887,
+   "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
+  },
+  "uhead": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+   "bytes": 9369632,
+   "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
+  },
+  "prm": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/prm_selector.json",
+   "bytes": 9360982,
+   "sha256": "0be50791aa0d0af5d2f2d239f8fdf551919f716c2c84b295553de06be28a3fcd"
+  },
+  "tier:think16k": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
+   "bytes": 26666,
+   "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
+  }
+ },
+ "finished_at": "2026-10-25T14:12:31+04:00",
+ "wall_secs": 18.6
+}

analysis/bayes_sc_2026-10-25/v8/bc8_usd_cheapest.arrays.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72d00641ff88ba44660e952983378e9b60bdbe41fa72386e90da50d770c08d73
+size 21885

analysis/bayes_sc_2026-10-25/v8/bc8_usd_cheapest.json

@@ -0,0 +1,142 @@
+{
+ "cost_model": {
+  "currency": "usd",
+  "unit": "USD",
+  "per_equiv": 2.0000000000000002e-07,
+  "scenario": "cheapest",
+  "base": "qwen3-8b:nothink"
+ },
+ "ref": {
+  "tier": "think16k",
+  "acc": 0.9279941219691403,
+  "equiv": 4668.220714810538
+ },
+ "frontier": [
+  {
+   "R": 2000,
+   "acc": 0.7883908890521675,
+   "equiv": 1129.5194709772227
+  },
+  {
+   "R": 4000,
+   "acc": 0.7883908890521675,
+   "equiv": 1129.5194709772227
+  },
+  {
+   "R": 6000,
+   "acc": 0.7883908890521675,
+   "equiv": 1129.5194709772227
+  },
+  {
+   "R": 8000,
+   "acc": 0.7883908890521675,
+   "equiv": 1129.5194709772227
+  },
+  {
+   "R": 12000,
+   "acc": 0.831373989713446,
+   "equiv": 2256.5245617718065
+  },
+  {
+   "R": 16000,
+   "acc": 0.8714705573632832,
+   "equiv": 3156.445024666737
+  },
+  {
+   "R": 24000,
+   "acc": 0.9091529337671879,
+   "equiv": 3961.4649679857243
+  },
+  {
+   "R": 32000,
+   "acc": 0.9218536790175291,
+   "equiv": 4277.911934501942
+  },
+  {
+   "R": 48000,
+   "acc": 0.9362338616563451,
+   "equiv": 4706.574734963788
+  }
+ ],
+ "matched": [
+  {
+   "seed": 0,
+   "acc": 0.9362338616563451,
+   "equiv": 4706.6886742941115,
+   "R": [
+    48781.018757585625,
+    47737.196817866185
+   ],
+   "delta_vs_agree2": [
+    0.008239739687204787,
+    0.0023617088275427736,
+    0.014697701270074514
+   ],
+   "actions": {
+    "mid:think16k": 475
+   }
+  },
+  {
+   "seed": 1,
+   "acc": 0.9318778209299884,
+   "equiv": 4678.954576466884,
+   "R": [
+    50114.318176677625,
+    50034.462920524595
+   ],
+   "delta_vs_agree2": [
+    0.0038836989608481164,
+    -0.0025204681431720364,
+    0.010970137503936177
+   ],
+   "actions": {
+    "mid:think16k": 463
+   }
+  },
+  {
+   "seed": 2,
+   "acc": 0.9338196704104126,
+   "equiv": 4655.15776214968,
+   "R": [
+    49992.40866974007,
+    50651.911602932625
+   ],
+   "delta_vs_agree2": [
+    0.005825548441272175,
+    -0.000526136244358139,
+    0.012229715545292322
+   ],
+   "actions": {
+    "mid:think16k": 443
+   }
+  },
+  {
+   "seed": 3,
+   "acc": 0.9354466253804977,
+   "equiv": 4696.135798257584,
+   "R": [
+    48932.595912206365,
+    51974.5228064271
+   ],
+   "delta_vs_agree2": [
+    0.007452503411357195,
+    0.0012595780413561463,
+    0.013699223260207826
+   ],
+   "actions": {
+    "mid:think16k": 463
+   }
+  }
+ ],
+ "agree2_by_tier": {
+  "think16k": {
+   "acc": 0.9279941219691403,
+   "equiv": 4668.220714810538,
+   "delta_vs_ref": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  }
+ }
+}

analysis/bayes_sc_2026-10-25/v8/bc8_usd_cheapest.log

@@ -0,0 +1,17 @@
+2026-10-25 14:12:14,966 online_rl_train INFO: ThinkBooster math evaluator loaded
+reference agree2 -> think16k: acc 0.928 equiv 4668 | base 0.788 at 1130 | always think16k 0.948 at 6346
+cost model usd: 1 equiv = 2e-07 USD
+agree2 -> think16k       acc 0.928 equiv   4668 (   +0) | vs reference +0.0 [+0.0, +0.0]
+R    2000: acc 0.788 equiv   1130
+R    4000: acc 0.788 equiv   1130
+R    6000: acc 0.788 equiv   1130
+R    8000: acc 0.788 equiv   1130
+R   12000: acc 0.831 equiv   2257
+R   16000: acc 0.871 equiv   3156
+R   24000: acc 0.909 equiv   3961
+R   32000: acc 0.922 equiv   4278
+R   48000: acc 0.936 equiv   4707
+parity: Bayes at agree2's cost acc 0.936 equiv 4707 (+38) | vs agree2 +0.8 [+0.2, +1.5] | R [48781, 47737] | actions {'mid:think16k': 475}
+split 1: Bayes at agree2's cost acc 0.932 equiv 4679 (+11) | vs agree2 +0.4 [-0.3, +1.1] | R [50114, 50034] | actions {'mid:think16k': 463}
+split 2: Bayes at agree2's cost acc 0.934 equiv 4655 (-13) | vs agree2 +0.6 [-0.1, +1.2] | R [49992, 50652] | actions {'mid:think16k': 443}
+split 3: Bayes at agree2's cost acc 0.935 equiv 4696 (+28) | vs agree2 +0.7 [+0.1, +1.4] | R [48933, 51975] | actions {'mid:think16k': 463}

analysis/bayes_sc_2026-10-25/v8/bc8_usd_cheapest.provenance.json

@@ -0,0 +1,100 @@
+{
+ "started_at": "2026-10-25T14:12:12+04:00",
+ "host": "ws006600",
+ "user": "vlad.smirnov",
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
+ "argv": [
+  "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_control.py",
+  "--pool-dir",
+  "workdir/selector/pool_mth_k8_chat",
+  "--uhead",
+  "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+  "--prm",
+  "workdir/selector/pool_mth_k8_chat/prm_selector.json",
+  "--prm-reads",
+  "workdir/selector/pool_mth_k8_chat/prm_reads.json",
+  "--step-tokens",
+  "<redacted>",
+  "--critics",
+  "agree",
+  "--task-prior",
+  "workdir/bayes/mth_task_prior.json",
+  "--tbins",
+  "3",
+  "--cost-model",
+  "usd",
+  "--price-scenario",
+  "cheapest",
+  "--tier",
+  "think16k=workdir/layer_probe/think16k_tier.json@qwen3-8b:think",
+  "--out",
+  "workdir/bayes/v8/bc8_usd_cheapest.json"
+ ],
+ "git": {
+  "commit": "f935c73993aae0c1137fdec2866aef2005007331",
+  "branch": "HEAD",
+  "subject": "Per-problem arrays in the Bayesian controllers; paired bootstrap against the hull mixture in currency_table",
+  "dirty": false,
+  "dirty_files": []
+ },
+ "gpus": {
+  "visible": "",
+  "devices": [
+   {
+    "index": "0",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
+   },
+   {
+    "index": "1",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
+   }
+  ]
+ },
+ "packages": {
+  "python": "3.11.15",
+  "torch": "2.9.1",
+  "vllm": "0.15.1",
+  "transformers": "4.57.6",
+  "speculators": "0.4.0.dev49",
+  "peft": null,
+  "accelerate": "1.13.0",
+  "numpy": "2.2.6",
+  "huggingface_hub": "0.36.2",
+  "openai": "2.32.0",
+  "wandb": "0.25.1"
+ },
+ "env": {
+  "CUDA_VISIBLE_DEVICES": ""
+ },
+ "models": [],
+ "files": {
+  "pool": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
+   "bytes": 65136887,
+   "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
+  },
+  "uhead": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+   "bytes": 9369632,
+   "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
+  },
+  "prm": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/prm_selector.json",
+   "bytes": 9360982,
+   "sha256": "0be50791aa0d0af5d2f2d239f8fdf551919f716c2c84b295553de06be28a3fcd"
+  },
+  "tier:think16k": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
+   "bytes": 26666,
+   "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
+  }
+ },
+ "finished_at": "2026-10-25T14:12:31+04:00",
+ "wall_secs": 18.9
+}

analysis/bayes_sc_2026-10-25/v8/bc8_usd_dear_32b.arrays.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72d00641ff88ba44660e952983378e9b60bdbe41fa72386e90da50d770c08d73
+size 21885

analysis/bayes_sc_2026-10-25/v8/bc8_usd_dear_32b.json

@@ -0,0 +1,142 @@
+{
+ "cost_model": {
+  "currency": "usd",
+  "unit": "USD",
+  "per_equiv": 2.0000000000000002e-07,
+  "scenario": "dear_32b",
+  "base": "qwen3-8b:nothink"
+ },
+ "ref": {
+  "tier": "think16k",
+  "acc": 0.9279941219691403,
+  "equiv": 4668.220714810538
+ },
+ "frontier": [
+  {
+   "R": 2000,
+   "acc": 0.7883908890521675,
+   "equiv": 1129.5194709772227
+  },
+  {
+   "R": 4000,
+   "acc": 0.7883908890521675,
+   "equiv": 1129.5194709772227
+  },
+  {
+   "R": 6000,
+   "acc": 0.7883908890521675,
+   "equiv": 1129.5194709772227
+  },
+  {
+   "R": 8000,
+   "acc": 0.7883908890521675,
+   "equiv": 1129.5194709772227
+  },
+  {
+   "R": 12000,
+   "acc": 0.831373989713446,
+   "equiv": 2256.5245617718065
+  },
+  {
+   "R": 16000,
+   "acc": 0.8714705573632832,
+   "equiv": 3156.445024666737
+  },
+  {
+   "R": 24000,
+   "acc": 0.9091529337671879,
+   "equiv": 3961.4649679857243
+  },
+  {
+   "R": 32000,
+   "acc": 0.9218536790175291,
+   "equiv": 4277.911934501942
+  },
+  {
+   "R": 48000,
+   "acc": 0.9362338616563451,
+   "equiv": 4706.574734963788
+  }
+ ],
+ "matched": [
+  {
+   "seed": 0,
+   "acc": 0.9362338616563451,
+   "equiv": 4706.6886742941115,
+   "R": [
+    48781.018757585625,
+    47737.196817866185
+   ],
+   "delta_vs_agree2": [
+    0.008239739687204787,
+    0.0023617088275427736,
+    0.014697701270074514
+   ],
+   "actions": {
+    "mid:think16k": 475
+   }
+  },
+  {
+   "seed": 1,
+   "acc": 0.9318778209299884,
+   "equiv": 4678.954576466884,
+   "R": [
+    50114.318176677625,
+    50034.462920524595
+   ],
+   "delta_vs_agree2": [
+    0.0038836989608481164,
+    -0.0025204681431720364,
+    0.010970137503936177
+   ],
+   "actions": {
+    "mid:think16k": 463
+   }
+  },
+  {
+   "seed": 2,
+   "acc": 0.9338196704104126,
+   "equiv": 4655.15776214968,
+   "R": [
+    49992.40866974007,
+    50651.911602932625
+   ],
+   "delta_vs_agree2": [
+    0.005825548441272175,
+    -0.000526136244358139,
+    0.012229715545292322
+   ],
+   "actions": {
+    "mid:think16k": 443
+   }
+  },
+  {
+   "seed": 3,
+   "acc": 0.9354466253804977,
+   "equiv": 4696.135798257584,
+   "R": [
+    48932.595912206365,
+    51974.5228064271
+   ],
+   "delta_vs_agree2": [
+    0.007452503411357195,
+    0.0012595780413561463,
+    0.013699223260207826
+   ],
+   "actions": {
+    "mid:think16k": 463
+   }
+  }
+ ],
+ "agree2_by_tier": {
+  "think16k": {
+   "acc": 0.9279941219691403,
+   "equiv": 4668.220714810538,
+   "delta_vs_ref": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  }
+ }
+}

analysis/bayes_sc_2026-10-25/v8/bc8_usd_dear_32b.log

@@ -0,0 +1,17 @@
+2026-10-25 14:12:15,010 online_rl_train INFO: ThinkBooster math evaluator loaded
+reference agree2 -> think16k: acc 0.928 equiv 4668 | base 0.788 at 1130 | always think16k 0.948 at 6346
+cost model usd: 1 equiv = 2e-07 USD
+agree2 -> think16k       acc 0.928 equiv   4668 (   +0) | vs reference +0.0 [+0.0, +0.0]
+R    2000: acc 0.788 equiv   1130
+R    4000: acc 0.788 equiv   1130
+R    6000: acc 0.788 equiv   1130
+R    8000: acc 0.788 equiv   1130
+R   12000: acc 0.831 equiv   2257
+R   16000: acc 0.871 equiv   3156
+R   24000: acc 0.909 equiv   3961
+R   32000: acc 0.922 equiv   4278
+R   48000: acc 0.936 equiv   4707
+parity: Bayes at agree2's cost acc 0.936 equiv 4707 (+38) | vs agree2 +0.8 [+0.2, +1.5] | R [48781, 47737] | actions {'mid:think16k': 475}
+split 1: Bayes at agree2's cost acc 0.932 equiv 4679 (+11) | vs agree2 +0.4 [-0.3, +1.1] | R [50114, 50034] | actions {'mid:think16k': 463}
+split 2: Bayes at agree2's cost acc 0.934 equiv 4655 (-13) | vs agree2 +0.6 [-0.1, +1.2] | R [49992, 50652] | actions {'mid:think16k': 443}
+split 3: Bayes at agree2's cost acc 0.935 equiv 4696 (+28) | vs agree2 +0.7 [+0.1, +1.4] | R [48933, 51975] | actions {'mid:think16k': 463}

analysis/bayes_sc_2026-10-25/v8/bc8_usd_dear_32b.provenance.json

@@ -0,0 +1,100 @@
+{
+ "started_at": "2026-10-25T14:12:12+04:00",
+ "host": "ws006600",
+ "user": "vlad.smirnov",
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
+ "argv": [
+  "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_control.py",
+  "--pool-dir",
+  "workdir/selector/pool_mth_k8_chat",
+  "--uhead",
+  "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+  "--prm",
+  "workdir/selector/pool_mth_k8_chat/prm_selector.json",
+  "--prm-reads",
+  "workdir/selector/pool_mth_k8_chat/prm_reads.json",
+  "--step-tokens",
+  "<redacted>",
+  "--critics",
+  "agree",
+  "--task-prior",
+  "workdir/bayes/mth_task_prior.json",
+  "--tbins",
+  "3",
+  "--cost-model",
+  "usd",
+  "--price-scenario",
+  "dear_32b",
+  "--tier",
+  "think16k=workdir/layer_probe/think16k_tier.json@qwen3-8b:think",
+  "--out",
+  "workdir/bayes/v8/bc8_usd_dear_32b.json"
+ ],
+ "git": {
+  "commit": "f935c73993aae0c1137fdec2866aef2005007331",
+  "branch": "HEAD",
+  "subject": "Per-problem arrays in the Bayesian controllers; paired bootstrap against the hull mixture in currency_table",
+  "dirty": false,
+  "dirty_files": []
+ },
+ "gpus": {
+  "visible": "",
+  "devices": [
+   {
+    "index": "0",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
+   },
+   {
+    "index": "1",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
+   }
+  ]
+ },
+ "packages": {
+  "python": "3.11.15",
+  "torch": "2.9.1",
+  "vllm": "0.15.1",
+  "transformers": "4.57.6",
+  "speculators": "0.4.0.dev49",
+  "peft": null,
+  "accelerate": "1.13.0",
+  "numpy": "2.2.6",
+  "huggingface_hub": "0.36.2",
+  "openai": "2.32.0",
+  "wandb": "0.25.1"
+ },
+ "env": {
+  "CUDA_VISIBLE_DEVICES": ""
+ },
+ "models": [],
+ "files": {
+  "pool": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
+   "bytes": 65136887,
+   "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
+  },
+  "uhead": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+   "bytes": 9369632,
+   "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
+  },
+  "prm": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/prm_selector.json",
+   "bytes": 9360982,
+   "sha256": "0be50791aa0d0af5d2f2d239f8fdf551919f716c2c84b295553de06be28a3fcd"
+  },
+  "tier:think16k": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
+   "bytes": 26666,
+   "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
+  }
+ },
+ "finished_at": "2026-10-25T14:12:31+04:00",
+ "wall_secs": 18.6
+}

analysis/bayes_sc_2026-10-25/v8/bc8_usd_openrouter_only.arrays.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35f44d70ba65998832380bcec974a2b88e323c0b5bee86dc0fec50b0a4a630e0
+size 27376

analysis/bayes_sc_2026-10-25/v8/bc8_usd_openrouter_only.json

@@ -0,0 +1,142 @@
+{
+ "cost_model": {
+  "currency": "usd",
+  "unit": "USD",
+  "per_equiv": 4.5500000000000004e-07,
+  "scenario": "openrouter_only",
+  "base": "qwen3-8b:nothink"
+ },
+ "ref": {
+  "tier": "think16k",
+  "acc": 0.9279941219691403,
+  "equiv": 4669.85060504731
+ },
+ "frontier": [
+  {
+   "R": 2000,
+   "acc": 0.7883908890521675,
+   "equiv": 1130.2242048913613
+  },
+  {
+   "R": 4000,
+   "acc": 0.7883908890521675,
+   "equiv": 1130.2242048913613
+  },
+  {
+   "R": 6000,
+   "acc": 0.7883908890521675,
+   "equiv": 1130.2242048913613
+  },
+  {
+   "R": 8000,
+   "acc": 0.7883908890521675,
+   "equiv": 1130.2242048913613
+  },
+  {
+   "R": 12000,
+   "acc": 0.831373989713446,
+   "equiv": 2257.3525633912636
+  },
+  {
+   "R": 16000,
+   "acc": 0.8714705573632832,
+   "equiv": 3157.4179504865865
+  },
+  {
+   "R": 24000,
+   "acc": 0.9091529337671879,
+   "equiv": 3962.602664607356
+  },
+  {
+   "R": 32000,
+   "acc": 0.9218536790175291,
+   "equiv": 4279.105650107214
+  },
+  {
+   "R": 48000,
+   "acc": 0.9362338616563451,
+   "equiv": 4707.833017439158
+  }
+ ],
+ "matched": [
+  {
+   "seed": 0,
+   "acc": 0.9362338616563451,
+   "equiv": 4704.38500952181,
+   "R": [
+    48776.77570445791,
+    47756.578806575955
+   ],
+   "delta_vs_agree2": [
+    0.008239739687204787,
+    0.0023617088275427736,
+    0.014697701270074514
+   ],
+   "actions": {
+    "mid:think16k": 475
+   }
+  },
+  {
+   "seed": 1,
+   "acc": 0.9318778209299884,
+   "equiv": 4689.905929013781,
+   "R": [
+    50112.86512598025,
+    50031.56149195534
+   ],
+   "delta_vs_agree2": [
+    0.0038836989608481164,
+    -0.0025204681431720364,
+    0.010970137503936177
+   ],
+   "actions": {
+    "mid:think16k": 467
+   }
+  },
+  {
+   "seed": 2,
+   "acc": 0.9338196704104126,
+   "equiv": 4656.432901977837,
+   "R": [
+    49992.40866974007,
+    50648.97436937731
+   ],
+   "delta_vs_agree2": [
+    0.005825548441272175,
+    -0.000526136244358139,
+    0.012229715545292322
+   ],
+   "actions": {
+    "mid:think16k": 443
+   }
+  },
+  {
+   "seed": 3,
+   "acc": 0.9354466253804977,
+   "equiv": 4684.456896939525,
+   "R": [
+    48931.177125216906,
+    51997.132852835726
+   ],
+   "delta_vs_agree2": [
+    0.007452503411357195,
+    0.0012595780413561463,
+    0.013699223260207826
+   ],
+   "actions": {
+    "mid:think16k": 460
+   }
+  }
+ ],
+ "agree2_by_tier": {
+  "think16k": {
+   "acc": 0.9279941219691403,
+   "equiv": 4669.85060504731,
+   "delta_vs_ref": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  }
+ }
+}

analysis/bayes_sc_2026-10-25/v8/bc8_usd_openrouter_only.log

@@ -0,0 +1,17 @@
+2026-10-25 14:12:14,989 online_rl_train INFO: ThinkBooster math evaluator loaded
+reference agree2 -> think16k: acc 0.928 equiv 4670 | base 0.788 at 1130 | always think16k 0.948 at 6347
+cost model usd: 1 equiv = 4.55e-07 USD
+agree2 -> think16k       acc 0.928 equiv   4670 (   +0) | vs reference +0.0 [+0.0, +0.0]
+R    2000: acc 0.788 equiv   1130
+R    4000: acc 0.788 equiv   1130
+R    6000: acc 0.788 equiv   1130
+R    8000: acc 0.788 equiv   1130
+R   12000: acc 0.831 equiv   2257
+R   16000: acc 0.871 equiv   3157
+R   24000: acc 0.909 equiv   3963
+R   32000: acc 0.922 equiv   4279
+R   48000: acc 0.936 equiv   4708
+parity: Bayes at agree2's cost acc 0.936 equiv 4704 (+35) | vs agree2 +0.8 [+0.2, +1.5] | R [48777, 47757] | actions {'mid:think16k': 475}
+split 1: Bayes at agree2's cost acc 0.932 equiv 4690 (+20) | vs agree2 +0.4 [-0.3, +1.1] | R [50113, 50032] | actions {'mid:think16k': 467}
+split 2: Bayes at agree2's cost acc 0.934 equiv 4656 (-13) | vs agree2 +0.6 [-0.1, +1.2] | R [49992, 50649] | actions {'mid:think16k': 443}
+split 3: Bayes at agree2's cost acc 0.935 equiv 4684 (+15) | vs agree2 +0.7 [+0.1, +1.4] | R [48931, 51997] | actions {'mid:think16k': 460}

analysis/bayes_sc_2026-10-25/v8/bc8_usd_openrouter_only.provenance.json

@@ -0,0 +1,100 @@
+{
+ "started_at": "2026-10-25T14:12:12+04:00",
+ "host": "ws006600",
+ "user": "vlad.smirnov",
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
+ "argv": [
+  "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_control.py",
+  "--pool-dir",
+  "workdir/selector/pool_mth_k8_chat",
+  "--uhead",
+  "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+  "--prm",
+  "workdir/selector/pool_mth_k8_chat/prm_selector.json",
+  "--prm-reads",
+  "workdir/selector/pool_mth_k8_chat/prm_reads.json",
+  "--step-tokens",
+  "<redacted>",
+  "--critics",
+  "agree",
+  "--task-prior",
+  "workdir/bayes/mth_task_prior.json",
+  "--tbins",
+  "3",
+  "--cost-model",
+  "usd",
+  "--price-scenario",
+  "openrouter_only",
+  "--tier",
+  "think16k=workdir/layer_probe/think16k_tier.json@qwen3-8b:think",
+  "--out",
+  "workdir/bayes/v8/bc8_usd_openrouter_only.json"
+ ],
+ "git": {
+  "commit": "f935c73993aae0c1137fdec2866aef2005007331",
+  "branch": "HEAD",
+  "subject": "Per-problem arrays in the Bayesian controllers; paired bootstrap against the hull mixture in currency_table",
+  "dirty": false,
+  "dirty_files": []
+ },
+ "gpus": {
+  "visible": "",
+  "devices": [
+   {
+    "index": "0",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
+   },
+   {
+    "index": "1",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
+   }
+  ]
+ },
+ "packages": {
+  "python": "3.11.15",
+  "torch": "2.9.1",
+  "vllm": "0.15.1",
+  "transformers": "4.57.6",
+  "speculators": "0.4.0.dev49",
+  "peft": null,
+  "accelerate": "1.13.0",
+  "numpy": "2.2.6",
+  "huggingface_hub": "0.36.2",
+  "openai": "2.32.0",
+  "wandb": "0.25.1"
+ },
+ "env": {
+  "CUDA_VISIBLE_DEVICES": ""
+ },
+ "models": [],
+ "files": {
+  "pool": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
+   "bytes": 65136887,
+   "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
+  },
+  "uhead": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+   "bytes": 9369632,
+   "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
+  },
+  "prm": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/prm_selector.json",
+   "bytes": 9360982,
+   "sha256": "0be50791aa0d0af5d2f2d239f8fdf551919f716c2c84b295553de06be28a3fcd"
+  },
+  "tier:think16k": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
+   "bytes": 26666,
+   "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
+  }
+ },
+ "finished_at": "2026-10-25T14:12:31+04:00",
+ "wall_secs": 18.4
+}

analysis/bayes_sc_2026-10-25/v8/currency_table.json

@@ -0,0 +1,1287 @@
+{
+ "tokens": {
+  "unit": "base tokens",
+  "per_equiv": 1.0,
+  "hull": [
+   {
+    "cost": 1124.9408523144746,
+    "acc": 0.7907788390889052,
+    "policy": "scboth:SC@1"
+   },
+   {
+    "cost": 4283.8179463629685,
+    "acc": 0.9405767817781043,
+    "policy": "sc32:agree2"
+   },
+   {
+    "cost": 5317.299044819985,
+    "acc": 0.9603232916972814,
+    "policy": "sc32:always think"
+   }
+  ],
+  "controllers": {
+   "bc8": {
+    "acc_minus_hull": [
+     [
+      -0.01154593444489512,
+      -0.021689004533106276,
+      -0.0011374065314038427
+     ],
+     [
+      -0.016537690823465616,
+      -0.027227445437807935,
+      -0.0057864715503241106
+     ],
+     [
+      -0.014129673297373455,
+      -0.024669886454969016,
+      -0.0034380419597849097
+     ],
+     [
+      -0.012066011997219907,
+      -0.022183695083823802,
+      -0.0015022208890707429
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      0.11104680758615171,
+      0.07828189469835561,
+      0.14699484475552319
+     ],
+     [
+      0.13607321514055926,
+      0.10065338632691762,
+      0.17507859415861599
+     ],
+     [
+      0.11810433862788394,
+      0.08327523872479879,
+      0.1535466443546841
+     ],
+     [
+      0.10847382209079726,
+      0.07328610723408384,
+      0.14558199254682896
+     ]
+    ]
+   },
+   "sc32": {
+    "acc_minus_hull": [
+     [
+      0.01727890836704822,
+      0.011744912763194603,
+      0.023328553095136276
+     ],
+     [
+      0.01272858047671425,
+      0.007354553991362199,
+      0.01836762727574271
+     ],
+     [
+      0.014321840159119404,
+      0.008847538807349205,
+      0.020179350935957775
+     ],
+     [
+      0.015851101629264938,
+      0.010512041070992366,
+      0.021600229987031472
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      -0.14411166729135594,
+      -0.16652181663998855,
+      -0.11978340541668497
+     ],
+     [
+      -0.13656536882693682,
+      -0.16006792100782916,
+      -0.11255592836424118
+     ],
+     [
+      -0.13812488260488387,
+      -0.16274373031985748,
+      -0.11228978008622358
+     ],
+     [
+      -0.14773946465099874,
+      -0.17138060336285513,
+      -0.12408898797752163
+     ]
+    ]
+   },
+   "sc8": {
+    "acc_minus_hull": [
+     [
+      -0.011831013057281568,
+      -0.020924259595917337,
+      -0.0026943518330644727
+     ],
+     [
+      -0.014340175221638775,
+      -0.02380221186441814,
+      -0.005281672300869335
+     ],
+     [
+      -0.013505554275997738,
+      -0.022774420917686774,
+      -0.004476986827302682
+     ],
+     [
+      -0.011598439360796067,
+      -0.020457417145376758,
+      -0.002429287859219795
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      0.08470615401783421,
+      0.052913132222845445,
+      0.11952293446691922
+     ],
+     [
+      0.10305505418229433,
+      0.06846897513455145,
+      0.14043684090919975
+     ],
+     [
+      0.0836006068515962,
+      0.051014442597240126,
+      0.11869868317743142
+     ],
+     [
+      0.08034059961749684,
+      0.04859655962100912,
+      0.11472260159294469
+     ]
+    ]
+   },
+   "scboth": {
+    "acc_minus_hull": [
+     [
+      0.011381752078486577,
+      0.006872236326157534,
+      0.016268995972369034
+     ],
+     [
+      0.00909360071205127,
+      0.004674066807279793,
+      0.01368502934783614
+     ],
+     [
+      0.0095842394268741,
+      0.005362710598833014,
+      0.014148305465628867
+     ],
+     [
+      0.008366983981449715,
+      0.004355469257275526,
+      0.012553652816093518
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      -0.14431008697788583,
+      -0.17160719113740078,
+      -0.11579903905198777
+     ],
+     [
+      -0.1413480024591749,
+      -0.16755285404347794,
+      -0.11433735732866102
+     ],
+     [
+      -0.14260974935373127,
+      -0.1688997492213235,
+      -0.11537052361907747
+     ],
+     [
+      -0.14828294930936392,
+      -0.1749079374749139,
+      -0.12123062275791205
+     ]
+    ]
+   },
+   "scbud": {
+    "acc_minus_hull": [
+     [
+      -0.012763604027751434,
+      -0.022245581387111427,
+      -0.003001987229314552
+     ],
+     [
+      -0.013413936019365491,
+      -0.02257638835839211,
+      -0.0038743259030610527
+     ],
+     [
+      -0.016406413180826942,
+      -0.025928803882051342,
+      -0.006839534464497963
+     ],
+     [
+      -0.012481679714065406,
+      -0.02199072499730993,
+      -0.003176210068166959
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      0.08697793061074388,
+      0.054973826539053486,
+      0.12306320985625598
+     ],
+     [
+      0.0658223665122728,
+      0.03310762138813217,
+      0.10214061336565375
+     ],
+     [
+      0.08179331943576984,
+      0.04991358266894612,
+      0.11812508365989428
+     ],
+     [
+      0.066359410448106,
+      0.03170757660972136,
+      0.1054058897793177
+     ]
+    ]
+   }
+  }
+ },
+ "gpu_s": {
+  "unit": "GPU-seconds",
+  "per_equiv": 0.0006341154090044388,
+  "hull": [
+   {
+    "cost": 0.7290160193332379,
+    "acc": 0.7907788390889052,
+    "policy": "scboth:SC@1"
+   },
+   {
+    "cost": 3.4543134172303263,
+    "acc": 0.929922850844967,
+    "policy": "scboth:agree2"
+   },
+   {
+    "cost": 5.253141469447148,
+    "acc": 0.9478324761204996,
+    "policy": "scboth:always think"
+   },
+   {
+    "cost": 13.76095541375684,
+    "acc": 0.9572005878030859,
+    "policy": "sc32:AC(0.95)+think@4"
+   },
+   {
+    "cost": 20.678831458294123,
+    "acc": 0.9603232916972814,
+    "policy": "sc32:always think"
+   }
+  ],
+  "controllers": {
+   "bc8": {
+    "acc_minus_hull": [
+     [
+      -0.00048154355904439473,
+      -0.007730431883272332,
+      0.006171427969939402
+     ],
+     [
+      0.00048195462864539864,
+      -0.007078243540458368,
+      0.008108301137940845
+     ],
+     [
+      0.002284477163780532,
+      -0.004887972860546388,
+      0.009277655654298604
+     ],
+     [
+      0.003591728521452854,
+      -0.003270988758254005,
+      0.010584169402185607
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      0.0032613595623585123,
+      -0.024845781466683484,
+      0.0351182697593119
+     ],
+     [
+      -0.002783022760997511,
+      -0.034224052848013865,
+      0.029567120765436414
+     ],
+     [
+      -0.03669478392009673,
+      -0.06359304530840613,
+      -0.0061231897176151715
+     ],
+     [
+      -0.0571685241224561,
+      -0.08379419009012114,
+      -0.02837579211958979
+     ]
+    ]
+   },
+   "sc32": {
+    "acc_minus_hull": [
+     [
+      -0.00760027306371133,
+      -0.014475690481403038,
+      -0.0008310043147514773
+     ],
+     [
+      -0.007782889804391352,
+      -0.015275805300946393,
+      -0.0006534257904663
+     ],
+     [
+      -0.009395025846732213,
+      -0.017491165745819545,
+      -0.0018847514330813963
+     ],
+     [
+      -0.009042885514006093,
+      -0.016713902476019588,
+      -0.001731084796915497
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      1.0735981614007977,
+      0.9415138019372935,
+      1.1988240443123437
+     ],
+     [
+      0.9239211703026753,
+      0.7907871694877033,
+      1.0513820772074665
+     ],
+     [
+      0.9767319261805385,
+      0.8434704144697326,
+      1.1047462004883983
+     ],
+     [
+      1.082060399531183,
+      0.9467845448576931,
+      1.2129171792397617
+     ]
+    ]
+   },
+   "sc8": {
+    "acc_minus_hull": [
+     [
+      -0.0016858458055734418,
+      -0.007151926978278105,
+      0.003893793354212172
+     ],
+     [
+      5.592856338770559e-05,
+      -0.0054508032955559995,
+      0.005653126710551691
+     ],
+     [
+      -0.002868407827873601,
+      -0.008516665206462699,
+      0.0030549398174657947
+     ],
+     [
+      -0.0018096210435353628,
+      -0.007758177437949999,
+      0.004416730344314157
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      0.01589699109308862,
+      -0.014653279134561844,
+      0.04950279856017859
+     ],
+     [
+      0.021314558716862164,
+      -0.011253524611173188,
+      0.05615939583292788
+     ],
+     [
+      0.013562872849064211,
+      -0.0177623173158587,
+      0.047546722784043274
+     ],
+     [
+      0.01039326387117967,
+      -0.020183755015209477,
+      0.04340494874304136
+     ]
+    ]
+   },
+   "scboth": {
+    "acc_minus_hull": [
+     [
+      -0.0016858458055734418,
+      -0.007151926978278105,
+      0.003893793354212172
+     ],
+     [
+      5.592856338770559e-05,
+      -0.0054508032955559995,
+      0.005653126710551691
+     ],
+     [
+      -0.002868407827873601,
+      -0.008516665206462699,
+      0.0030549398174657947
+     ],
+     [
+      -0.0018096210435353628,
+      -0.007758177437949999,
+      0.004416730344314157
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      0.01589699109308862,
+      -0.014653279134561844,
+      0.04950279856017859
+     ],
+     [
+      0.021314558716862164,
+      -0.011253524611173188,
+      0.05615939583292788
+     ],
+     [
+      0.013562872849064211,
+      -0.0177623173158587,
+      0.047546722784043274
+     ],
+     [
+      0.01039326387117967,
+      -0.020183755015209477,
+      0.04340494874304136
+     ]
+    ]
+   },
+   "scbud": {
+    "acc_minus_hull": [
+     [
+      -0.001975094951340463,
+      -0.008351364144323446,
+      0.004671268656209748
+     ],
+     [
+      0.00397995179486257,
+      -0.004830132768530077,
+      0.012795608565363424
+     ],
+     [
+      -0.0015036984082572224,
+      -0.00749028953020849,
+      0.004561025654270194
+     ],
+     [
+      -0.0015299221051515927,
+      -0.007715180846166277,
+      0.0044897653115898006
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      0.010390459635348348,
+      -0.020932530159203898,
+      0.04403831969254968
+     ],
+     [
+      0.012379209238011946,
+      -0.01880095002203893,
+      0.04707347028879613
+     ],
+     [
+      0.01976795325589742,
+      -0.011908985163736324,
+      0.05347997084323053
+     ],
+     [
+      0.003533757204437382,
+      -0.0291597789461571,
+      0.04147560355249579
+     ]
+    ]
+   }
+  }
+ },
+ "usd_cheapest": {
+  "unit": "USD",
+  "per_equiv": 2.0000000000000002e-07,
+  "hull": [
+   {
+    "cost": 0.00022992130786186627,
+    "acc": 0.7907788390889052,
+    "policy": "sc8:SC@1"
+   },
+   {
+    "cost": 0.0009421097997795738,
+    "acc": 0.929922850844967,
+    "policy": "sc8:agree2"
+   },
+   {
+    "cost": 0.0010301091219691403,
+    "acc": 0.9405767817781043,
+    "policy": "sc32:agree2"
+   },
+   {
+    "cost": 0.0014912116385011021,
+    "acc": 0.9603232916972814,
+    "policy": "sc32:always think"
+   }
+  ],
+  "controllers": {
+   "bc8": {
+    "acc_minus_hull": [
+     [
+      0.006461853195983721,
+      0.00023723940444089547,
+      0.012824841407981088
+     ],
+     [
+      0.0031895237842326334,
+      -0.003773076827493333,
+      0.010254499139280114
+     ],
+     [
+      0.0060612350213929776,
+      -0.0005033765276781783,
+      0.012814469917665779
+     ],
+     [
+      0.0060869710999496875,
+      -0.0004956335967164088,
+      0.012472244540277557
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      -0.0532063450098913,
+      -0.08032570572421102,
+      -0.022156298094874495
+     ],
+     [
+      -0.023445205834801475,
+      -0.05121628786760856,
+      0.007151708436422067
+     ],
+     [
+      -0.044406585382604336,
+      -0.07245505214595364,
+      -0.014705170323175008
+     ],
+     [
+      -0.04911023839032702,
+      -0.07746816257675937,
+      -0.018718227898915377
+     ]
+    ]
+   },
+   "sc32": {
+    "acc_minus_hull": [
+     [
+      0.00847387069304179,
+      0.002752211932109202,
+      0.0147067869511535
+     ],
+     [
+      0.00543196119618611,
+      -0.0007901506437870733,
+      0.012287642868848895
+     ],
+     [
+      0.00717665576660485,
+      0.0012001277069492444,
+      0.013546023587525744
+     ],
+     [
+      0.00899974445953801,
+      0.0032905861335968184,
+      0.015207017410604217
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      -0.05146112650207901,
+      -0.08056158073637068,
+      -0.020214674567855802
+     ],
+     [
+      -0.05149310055490697,
+      -0.08117172461135748,
+      -0.020903399342985136
+     ],
+     [
+      -0.04031006638550416,
+      -0.07139890154140159,
+      -0.007567716145226131
+     ],
+     [
+      -0.059042036571112066,
+      -0.08773568656607754,
+      -0.02924173278476002
+     ]
+    ]
+   },
+   "sc8": {
+    "acc_minus_hull": [
+     [
+      0.007576206175890477,
+      0.0025087311362865395,
+      0.013045562464691556
+     ],
+     [
+      0.003766617344983701,
+      -0.0013999603117203906,
+      0.009214176870289892
+     ],
+     [
+      0.004563308966810569,
+      -0.0004414767163326162,
+      0.009982355994823666
+     ],
+     [
+      0.007148919768599523,
+      0.0021915383390930733,
+      0.012599007002087456
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      -0.06543636536668762,
+      -0.09135087534327356,
+      -0.037435888101107796
+     ],
+     [
+      -0.04401704481886837,
+      -0.07284031632815699,
+      -0.014262590641537556
+     ],
+     [
+      -0.06380618549466399,
+      -0.08943566919345203,
+      -0.03604802468648296
+     ],
+     [
+      -0.06736791249423935,
+      -0.09356042224189386,
+      -0.03959883576869828
+     ]
+    ]
+   },
+   "scboth": {
+    "acc_minus_hull": [
+     [
+      0.005947591240978716,
+      -0.0025791257593269336,
+      0.014957702081186372
+     ],
+     [
+      -0.002584385097995378,
+      -0.009911241786641621,
+      0.005086326768490827
+     ],
+     [
+      0.004981732806636208,
+      -0.0029627403610281494,
+      0.013296864663134305
+     ],
+     [
+      0.008135585203617056,
+      -0.0001171314641442045,
+      0.0168435688710609
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      -0.04843912481007129,
+      -0.08004102905532201,
+      -0.01546400925358868
+     ],
+     [
+      0.0013963669178986837,
+      -0.033358295836115555,
+      0.037145039593570066
+     ],
+     [
+      -0.039138578334908636,
+      -0.07146530699156732,
+      -0.006454509662151425
+     ],
+     [
+      -0.039401684670443915,
+      -0.07101248647668243,
+      -0.007422798030194643
+     ]
+    ]
+   },
+   "scbud": {
+    "acc_minus_hull": [
+     [
+      0.007533829413569287,
+      0.001341474908743934,
+      0.013617621373855814
+     ],
+     [
+      0.004074135588326202,
+      -0.002015959278434806,
+      0.010195952000522916
+     ],
+     [
+      0.0024343497149646575,
+      -0.0035408088080407314,
+      0.0084421258722923
+     ],
+     [
+      0.006984600043258462,
+      0.0008704167515859184,
+      0.012992727213065788
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      -0.05720382831094939,
+      -0.08417090680475348,
+      -0.028265662067081195
+     ],
+     [
+      -0.07865556268370588,
+      -0.10500652075273233,
+      -0.049304273592419914
+     ],
+     [
+      -0.06516232712183578,
+      -0.0917846688165415,
+      -0.036465432932704704
+     ],
+     [
+      -0.07342046969120042,
+      -0.10118522293955877,
+      -0.04124073904680408
+     ]
+    ]
+   }
+  }
+ },
+ "usd_openrouter_only": {
+  "unit": "USD",
+  "per_equiv": 4.5500000000000004e-07,
+  "hull": [
+   {
+    "cost": 0.000523391629316679,
+    "acc": 0.7907788390889052,
+    "policy": "sc8:SC@1"
+   },
+   {
+    "cost": 0.0014912116385011026,
+    "acc": 0.9603232916972814,
+    "policy": "sc32:always think"
+   }
+  ],
+  "controllers": {
+   "bc8": {
+    "acc_minus_hull": [
+     [
+      -0.024089430040936284,
+      -0.035795633462789966,
+      -0.011859714495643965
+     ],
+     [
+      -0.02844547076729296,
+      -0.0403589797417865,
+      -0.015528235541093768
+     ],
+     [
+      -0.026503621286868898,
+      -0.03831216542458276,
+      -0.013800251915608285
+     ],
+     [
+      -0.024876666316783876,
+      -0.03679279941219691,
+      -0.012332056261152519
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      0.5812175652220934,
+      0.5243003483552899,
+      0.6372331591782454
+     ],
+     [
+      0.6058483956040768,
+      0.5476871666208383,
+      0.6624092921113377
+     ],
+     [
+      0.5811971521971844,
+      0.5232408833209826,
+      0.6382887495144551
+     ],
+     [
+      0.5797636910122075,
+      0.5195947649171093,
+      0.6381106633300871
+     ]
+    ]
+   },
+   "sc32": {
+    "acc_minus_hull": [
+     [
+      -0.002387950036737693,
+      -0.007255694342395298,
+      0.002387950036737693
+     ],
+     [
+      -0.002204261572373255,
+      -0.00615356355620867,
+      0.002204261572373255
+     ],
+     [
+      -0.0017450404114621601,
+      -0.00615356355620867,
+      0.002571638501102131
+     ],
+     [
+      -0.0033063923585598823,
+      -0.00707200587803086,
+      0.0002755326965466569
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      -0.15043938841044702,
+      -0.1744256228639952,
+      -0.12680480574523081
+     ],
+     [
+      -0.15858652125182837,
+      -0.18102165272068413,
+      -0.1370119177141602
+     ],
+     [
+      -0.15779677837051487,
+      -0.18188544779191038,
+      -0.13637997604005572
+     ],
+     [
+      -0.14605338799686296,
+      -0.16856229733308767,
+      -0.12426933629383971
+     ]
+    ]
+   },
+   "sc8": {
+    "acc_minus_hull": [
+     [
+      -0.024476487876561353,
+      -0.034993800514327704,
+      -0.013499954077883914
+     ],
+     [
+      -0.025165319617927995,
+      -0.036188923585598824,
+      -0.014234707935341666
+     ],
+     [
+      -0.025349008082292433,
+      -0.036095931300514326,
+      -0.014188785819250557
+     ],
+     [
+      -0.025303085966201323,
+      -0.035775624540778844,
+      -0.014096941587068338
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      0.5330814208554671,
+      0.47909890146732326,
+      0.5880885982065726
+     ],
+     [
+      0.5540628489231623,
+      0.4993388975245331,
+      0.6099124561612825
+     ],
+     [
+      0.5305390807229236,
+      0.47670982007855967,
+      0.5854206884726549
+     ],
+     [
+      0.5250235396658292,
+      0.472202835658319,
+      0.5776393799507212
+     ]
+    ]
+   },
+   "scboth": {
+    "acc_minus_hull": [
+     [
+      -0.002709404849375459,
+      -0.009047804922850844,
+      0.003719691403379868
+     ],
+     [
+      -0.002020573108008817,
+      -0.006612784717119765,
+      0.002755326965466569
+     ],
+     [
+      -0.0005510653930933138,
+      -0.0048677443056576045,
+      0.004132990448199853
+     ],
+     [
+      -0.0022501836884643643,
+      -0.006109937545922116,
+      0.0016072740631888317
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      -0.16955905817011319,
+      -0.19479604305419604,
+      -0.14498247995196106
+     ],
+     [
+      -0.16504706885748432,
+      -0.18956744646922988,
+      -0.14073116481625095
+     ],
+     [
+      -0.16789743150206637,
+      -0.19399132077984557,
+      -0.14440360324613755
+     ],
+     [
+      -0.17356432799292065,
+      -0.19917782540984438,
+      -0.15050376803231216
+     ]
+    ]
+   },
+   "scbud": {
+    "acc_minus_hull": [
+     [
+      -0.025716385011021307,
+      -0.03710966201322557,
+      -0.013868479059515063
+     ],
+     [
+      -0.024017266715650257,
+      -0.03518093313739897,
+      -0.012674504041146216
+     ],
+     [
+      -0.02902277736958119,
+      -0.040135929463629684,
+      -0.017266715650257163
+     ],
+     [
+      -0.026083761939750184,
+      -0.03692367744305658,
+      -0.014600936811168272
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      0.5266090507648056,
+      0.4757842841987392,
+      0.5791280535767499
+     ],
+     [
+      0.5020126128419893,
+      0.45163260264468635,
+      0.5535362852806782
+     ],
+     [
+      0.5196884944831792,
+      0.4685816649085815,
+      0.5712519835354636
+     ],
+     [
+      0.48981611738438446,
+      0.44149995600214287,
+      0.540441658966268
+     ]
+    ]
+   }
+  }
+ },
+ "usd_dear_32b": {
+  "unit": "USD",
+  "per_equiv": 2.0000000000000002e-07,
+  "hull": [
+   {
+    "cost": 0.00022992130786186627,
+    "acc": 0.7907788390889052,
+    "policy": "sc8:SC@1"
+   },
+   {
+    "cost": 0.0009421097997795738,
+    "acc": 0.929922850844967,
+    "policy": "sc8:agree2"
+   },
+   {
+    "cost": 0.0012692386480529024,
+    "acc": 0.9478324761204996,
+    "policy": "sc8:always think"
+   },
+   {
+    "cost": 0.002509960915686995,
+    "acc": 0.9572005878030859,
+    "policy": "sc32:AC(0.95)+think@4"
+   },
+   {
+    "cost": 0.0030334256869948566,
+    "acc": 0.9603232916972814,
+    "policy": "sc32:always think"
+   }
+  ],
+  "controllers": {
+   "bc8": {
+    "acc_minus_hull": [
+     [
+      0.006461853195983721,
+      0.00023723940444089547,
+      0.012824841407981088
+     ],
+     [
+      0.0031895237842326334,
+      -0.003773076827493333,
+      0.010254499139280114
+     ],
+     [
+      0.0060612350213929776,
+      -0.0005033765276781783,
+      0.012814469917665779
+     ],
+     [
+      0.0060869710999496875,
+      -0.0004956335967164088,
+      0.012472244540277557
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      -0.10974827000647636,
+      -0.13164471872184877,
+      -0.08750871203781087
+     ],
+     [
+      -0.042980839821979866,
+      -0.0677209485479761,
+      -0.015696041542626302
+     ],
+     [
+      -0.08117710959093671,
+      -0.10381168964096052,
+      -0.05735558323834158
+     ],
+     [
+      -0.09949844579497735,
+      -0.12349471610974982,
+      -0.07626208276844426
+     ]
+    ]
+   },
+   "sc32": {
+    "acc_minus_hull": [
+     [
+      -0.011346723989888419,
+      -0.019964106033219495,
+      -0.002960933184705086
+     ],
+     [
+      -0.010059606225570629,
+      -0.019071796604454058,
+      -0.0016658928326801245
+     ],
+     [
+      -0.008972906420464779,
+      -0.017919818285893632,
+      -0.00038094429517570525
+     ],
+     [
+      -0.009798900391291632,
+      -0.018426240185688145,
+      -0.001526370165657511
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      0.4743101864247081,
+      0.40188290308329594,
+      0.545478444534539
+     ],
+     [
+      0.4162609787291198,
+      0.3502323470544917,
+      0.4780863209115802
+     ],
+     [
+      0.4342922124038393,
+      0.3624034408695393,
+      0.5034776263996422
+     ],
+     [
+      0.43488201009475924,
+      0.36296893772890954,
+      0.5039562088447429
+     ]
+    ]
+   },
+   "sc8": {
+    "acc_minus_hull": [
+     [
+      0.007576206175890477,
+      0.0025087311362865395,
+      0.013045562464691556
+     ],
+     [
+      0.004520742121220974,
+      -0.00042030120219384076,
+      0.009825390864408734
+     ],
+     [
+      0.004563308966810569,
+      -0.0004414767163326162,
+      0.009982355994823666
+     ],
+     [
+      0.007148919768599523,
+      0.0021915383390930733,
+      0.012599007002087456
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      -0.08346617864261463,
+      -0.10673399083793,
+      -0.058709905582825926
+     ],
+     [
+      -0.04401704481886837,
+      -0.07284031632815699,
+      -0.014262590641537556
+     ],
+     [
+      -0.07453530543720732,
+      -0.09899771618904023,
+      -0.04877543454532804
+     ],
+     [
+      -0.07849186281833131,
+      -0.10345349813277324,
+      -0.05257902379449454
+     ]
+    ]
+   },
+   "scboth": {
+    "acc_minus_hull": [
+     [
+      0.007576206175890477,
+      0.0025087311362865395,
+      0.013045562464691556
+     ],
+     [
+      0.004520742121220974,
+      -0.00042030120219384076,
+      0.009825390864408734
+     ],
+     [
+      0.004563308966810569,
+      -0.0004414767163326162,
+      0.009982355994823666
+     ],
+     [
+      0.007148919768599523,
+      0.0021915383390930733,
+      0.012599007002087456
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      -0.08346617864261463,
+      -0.10673399083793,
+      -0.058709905582825926
+     ],
+     [
+      -0.04401704481886837,
+      -0.07284031632815699,
+      -0.014262590641537556
+     ],
+     [
+      -0.07453530543720732,
+      -0.09899771618904023,
+      -0.04877543454532804
+     ],
+     [
+      -0.07849186281833131,
+      -0.10345349813277324,
+      -0.05257902379449454
+     ]
+    ]
+   },
+   "scbud": {
+    "acc_minus_hull": [
+     [
+      0.007533829413569287,
+      0.001341474908743934,
+      0.013617621373855814
+     ],
+     [
+      0.004936518010892989,
+      -0.0010581317615102007,
+      0.010792463014637826
+     ],
+     [
+      0.0024343497149646575,
+      -0.0035408088080407314,
+      0.0084421258722923
+     ],
+     [
+      0.006984600043258462,
+      0.0008704167515859184,
+      0.012992727213065788
+     ]
+    ],
+    "cost_vs_hull": [
+     [
+      -0.05720382831094939,
+      -0.08417090680475348,
+      -0.028265662067081195
+     ],
+     [
+      -0.0831094764116378,
+      -0.10884936074165545,
+      -0.05485464715086791
+     ],
+     [
+      -0.06516232712183578,
+      -0.0917846688165415,
+      -0.036465432932704704
+     ],
+     [
+      -0.07342046969120042,
+      -0.10118522293955877,
+      -0.04124073904680408
+     ]
+    ]
+   }
+  }
+ }
+}

analysis/bayes_sc_2026-10-25/v8/currency_table.log

@@ -0,0 +1,55 @@
+== tokens (base tokens): hull SC@1 0.791@1125 -> agree2 0.941@4284 -> always think 0.960@5317
+  bc8     acc - hull: -1.2 [-2.2, -0.1] | -1.7 [-2.7, -0.6] | -1.4 [-2.5, -0.3] | -1.2 [-2.2, -0.2]
+          cost vs hull at its acc: +11.1% [+7.8, +14.7] | +13.6% [+10.1, +17.5] | +11.8% [+8.3, +15.4] | +10.8% [+7.3, +14.6]
+  sc32    acc - hull: +1.7 [+1.2, +2.3] | +1.3 [+0.7, +1.8] | +1.4 [+0.9, +2.0] | +1.6 [+1.1, +2.2]
+          cost vs hull at its acc: -14.4% [-16.7, -12.0] | -13.7% [-16.0, -11.3] | -13.8% [-16.3, -11.2] | -14.8% [-17.1, -12.4]
+  sc8     acc - hull: -1.2 [-2.1, -0.3] | -1.4 [-2.4, -0.5] | -1.4 [-2.3, -0.4] | -1.2 [-2.0, -0.2]
+          cost vs hull at its acc: +8.5% [+5.3, +12.0] | +10.3% [+6.8, +14.0] | +8.4% [+5.1, +11.9] | +8.0% [+4.9, +11.5]
+  scboth  acc - hull: +1.1 [+0.7, +1.6] | +0.9 [+0.5, +1.4] | +1.0 [+0.5, +1.4] | +0.8 [+0.4, +1.3]
+          cost vs hull at its acc: -14.4% [-17.2, -11.6] | -14.1% [-16.8, -11.4] | -14.3% [-16.9, -11.5] | -14.8% [-17.5, -12.1]
+  scbud   acc - hull: -1.3 [-2.2, -0.3] | -1.3 [-2.3, -0.4] | -1.6 [-2.6, -0.7] | -1.2 [-2.2, -0.3]
+          cost vs hull at its acc: +8.7% [+5.5, +12.3] | +6.6% [+3.3, +10.2] | +8.2% [+5.0, +11.8] | +6.6% [+3.2, +10.5]
+== gpu_s (GPU-seconds): hull SC@1 0.791@0.729 -> agree2 0.930@3.454 -> always think 0.948@5.253 -> AC(0.95)+think@4 0.957@13.76 -> always think 0.960@20.68
+  bc8     acc - hull: -0.0 [-0.8, +0.6] | +0.0 [-0.7, +0.8] | +0.2 [-0.5, +0.9] | +0.4 [-0.3, +1.1]
+          cost vs hull at its acc: +0.3% [-2.5, +3.5] | -0.3% [-3.4, +3.0] | -3.7% [-6.4, -0.6] | -5.7% [-8.4, -2.8]
+  sc32    acc - hull: -0.8 [-1.4, -0.1] | -0.8 [-1.5, -0.1] | -0.9 [-1.7, -0.2] | -0.9 [-1.7, -0.2]
+          cost vs hull at its acc: +107.4% [+94.2, +119.9] | +92.4% [+79.1, +105.1] | +97.7% [+84.3, +110.5] | +108.2% [+94.7, +121.3]
+  sc8     acc - hull: -0.2 [-0.7, +0.4] | +0.0 [-0.5, +0.6] | -0.3 [-0.9, +0.3] | -0.2 [-0.8, +0.4]
+          cost vs hull at its acc: +1.6% [-1.5, +5.0] | +2.1% [-1.1, +5.6] | +1.4% [-1.8, +4.8] | +1.0% [-2.0, +4.3]
+  scboth  acc - hull: -0.2 [-0.7, +0.4] | +0.0 [-0.5, +0.6] | -0.3 [-0.9, +0.3] | -0.2 [-0.8, +0.4]
+          cost vs hull at its acc: +1.6% [-1.5, +5.0] | +2.1% [-1.1, +5.6] | +1.4% [-1.8, +4.8] | +1.0% [-2.0, +4.3]
+  scbud   acc - hull: -0.2 [-0.8, +0.5] | +0.4 [-0.5, +1.3] | -0.2 [-0.7, +0.5] | -0.2 [-0.8, +0.4]
+          cost vs hull at its acc: +1.0% [-2.1, +4.4] | +1.2% [-1.9, +4.7] | +2.0% [-1.2, +5.3] | +0.4% [-2.9, +4.1]
+== usd_cheapest (USD): hull SC@1 0.791@0.0002299 -> agree2 0.930@0.0009421 -> agree2 0.941@0.00103 -> always think 0.960@0.001491
+  bc8     acc - hull: +0.6 [+0.0, +1.3] | +0.3 [-0.4, +1.0] | +0.6 [-0.1, +1.3] | +0.6 [-0.0, +1.2]
+          cost vs hull at its acc: -5.3% [-8.0, -2.2] | -2.3% [-5.1, +0.7] | -4.4% [-7.2, -1.5] | -4.9% [-7.7, -1.9]
+  sc32    acc - hull: +0.8 [+0.3, +1.5] | +0.5 [-0.1, +1.2] | +0.7 [+0.1, +1.4] | +0.9 [+0.3, +1.5]
+          cost vs hull at its acc: -5.1% [-8.1, -2.0] | -5.1% [-8.1, -2.1] | -4.0% [-7.1, -0.8] | -5.9% [-8.8, -2.9]
+  sc8     acc - hull: +0.8 [+0.3, +1.3] | +0.4 [-0.1, +0.9] | +0.5 [-0.0, +1.0] | +0.7 [+0.2, +1.3]
+          cost vs hull at its acc: -6.5% [-9.1, -3.7] | -4.4% [-7.3, -1.4] | -6.4% [-8.9, -3.6] | -6.7% [-9.4, -4.0]
+  scboth  acc - hull: +0.6 [-0.3, +1.5] | -0.3 [-1.0, +0.5] | +0.5 [-0.3, +1.3] | +0.8 [-0.0, +1.7]
+          cost vs hull at its acc: -4.8% [-8.0, -1.5] | +0.1% [-3.3, +3.7] | -3.9% [-7.1, -0.6] | -3.9% [-7.1, -0.7]
+  scbud   acc - hull: +0.8 [+0.1, +1.4] | +0.4 [-0.2, +1.0] | +0.2 [-0.4, +0.8] | +0.7 [+0.1, +1.3]
+          cost vs hull at its acc: -5.7% [-8.4, -2.8] | -7.9% [-10.5, -4.9] | -6.5% [-9.2, -3.6] | -7.3% [-10.1, -4.1]
+== usd_openrouter_only (USD): hull SC@1 0.791@0.0005234 -> always think 0.960@0.001491
+  bc8     acc - hull: -2.4 [-3.6, -1.2] | -2.8 [-4.0, -1.6] | -2.7 [-3.8, -1.4] | -2.5 [-3.7, -1.2]
+          cost vs hull at its acc: +58.1% [+52.4, +63.7] | +60.6% [+54.8, +66.2] | +58.1% [+52.3, +63.8] | +58.0% [+52.0, +63.8]
+  sc32    acc - hull: -0.2 [-0.7, +0.2] | -0.2 [-0.6, +0.2] | -0.2 [-0.6, +0.3] | -0.3 [-0.7, +0.0]
+          cost vs hull at its acc: -15.0% [-17.4, -12.7] | -15.9% [-18.1, -13.7] | -15.8% [-18.2, -13.6] | -14.6% [-16.9, -12.4]
+  sc8     acc - hull: -2.4 [-3.5, -1.3] | -2.5 [-3.6, -1.4] | -2.5 [-3.6, -1.4] | -2.5 [-3.6, -1.4]
+          cost vs hull at its acc: +53.3% [+47.9, +58.8] | +55.4% [+49.9, +61.0] | +53.1% [+47.7, +58.5] | +52.5% [+47.2, +57.8]
+  scboth  acc - hull: -0.3 [-0.9, +0.4] | -0.2 [-0.7, +0.3] | -0.1 [-0.5, +0.4] | -0.2 [-0.6, +0.2]
+          cost vs hull at its acc: -17.0% [-19.5, -14.5] | -16.5% [-19.0, -14.1] | -16.8% [-19.4, -14.4] | -17.4% [-19.9, -15.1]
+  scbud   acc - hull: -2.6 [-3.7, -1.4] | -2.4 [-3.5, -1.3] | -2.9 [-4.0, -1.7] | -2.6 [-3.7, -1.5]
+          cost vs hull at its acc: +52.7% [+47.6, +57.9] | +50.2% [+45.2, +55.4] | +52.0% [+46.9, +57.1] | +49.0% [+44.1, +54.0]
+== usd_dear_32b (USD): hull SC@1 0.791@0.0002299 -> agree2 0.930@0.0009421 -> always think 0.948@0.001269 -> AC(0.95)+think@4 0.957@0.00251 -> always think 0.960@0.003033
+  bc8     acc - hull: +0.6 [+0.0, +1.3] | +0.3 [-0.4, +1.0] | +0.6 [-0.1, +1.3] | +0.6 [-0.0, +1.2]
+          cost vs hull at its acc: -11.0% [-13.2, -8.8] | -4.3% [-6.8, -1.6] | -8.1% [-10.4, -5.7] | -9.9% [-12.3, -7.6]
+  sc32    acc - hull: -1.1 [-2.0, -0.3] | -1.0 [-1.9, -0.2] | -0.9 [-1.8, -0.0] | -1.0 [-1.8, -0.2]
+          cost vs hull at its acc: +47.4% [+40.2, +54.5] | +41.6% [+35.0, +47.8] | +43.4% [+36.2, +50.3] | +43.5% [+36.3, +50.4]
+  sc8     acc - hull: +0.8 [+0.3, +1.3] | +0.5 [-0.0, +1.0] | +0.5 [-0.0, +1.0] | +0.7 [+0.2, +1.3]
+          cost vs hull at its acc: -8.3% [-10.7, -5.9] | -4.4% [-7.3, -1.4] | -7.5% [-9.9, -4.9] | -7.8% [-10.3, -5.3]
+  scboth  acc - hull: +0.8 [+0.3, +1.3] | +0.5 [-0.0, +1.0] | +0.5 [-0.0, +1.0] | +0.7 [+0.2, +1.3]
+          cost vs hull at its acc: -8.3% [-10.7, -5.9] | -4.4% [-7.3, -1.4] | -7.5% [-9.9, -4.9] | -7.8% [-10.3, -5.3]
+  scbud   acc - hull: +0.8 [+0.1, +1.4] | +0.5 [-0.1, +1.1] | +0.2 [-0.4, +0.8] | +0.7 [+0.1, +1.3]
+          cost vs hull at its acc: -5.7% [-8.4, -2.8] | -8.3% [-10.9, -5.5] | -6.5% [-9.2, -3.6] | -7.3% [-10.1, -4.1]

analysis/bayes_sc_2026-10-25/v8/sc32_gpu_s.arrays.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b10e6addd10ff1db6b8ebd9780f33451dcc3526c8254a5d3e43c3fabd8cfe34
+size 325078

analysis/bayes_sc_2026-10-25/v8/sc32_gpu_s.json

@@ -0,0 +1,490 @@
+{
+ "ref": {
+  "acc": 0.9405767817781043,
+  "equiv": 14776.582078233669
+ },
+ "cost_model": {
+  "currency": "gpu_s",
+  "unit": "GPU-seconds",
+  "per_equiv": 0.0006341154090044388,
+  "scenario": null,
+  "base": "qwen3-8b:nothink"
+ },
+ "baselines": {
+  "agree2": {
+   "acc": 0.9405767817781043,
+   "equiv": 14776.582078233669,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "SC@1": {
+   "acc": 0.7907788390889052,
+   "equiv": 1149.6582624885161,
+   "delta_vs_agree2": [
+    -0.14979794268919913,
+    -0.16256429096252756,
+    -0.1369374540778839
+   ]
+  },
+  "SC@2": {
+   "acc": 0.8118111682586333,
+   "equiv": 2299.3165249770323,
+   "delta_vs_agree2": [
+    -0.12876561351947097,
+    -0.14226671565025717,
+    -0.11544819985304923
+   ]
+  },
+  "SC@3": {
+   "acc": 0.8444158706833211,
+   "equiv": 3448.9747874655477,
+   "delta_vs_agree2": [
+    -0.09616091109478325,
+    -0.10911324393828067,
+    -0.08357825128581925
+   ]
+  },
+  "SC@4": {
+   "acc": 0.8553453343130052,
+   "equiv": 4598.6330499540645,
+   "delta_vs_agree2": [
+    -0.0852314474650992,
+    -0.09809652828802351,
+    -0.07274063188831742
+   ]
+  },
+  "SC@5": {
+   "acc": 0.8630602498163116,
+   "equiv": 5748.29131244258,
+   "delta_vs_agree2": [
+    -0.0775165319617928,
+    -0.09056070903747244,
+    -0.06493387215282881
+   ]
+  },
+  "SC@6": {
+   "acc": 0.8693056576047025,
+   "equiv": 6897.949574931095,
+   "delta_vs_agree2": [
+    -0.07127112417340191,
+    -0.08440714548126378,
+    -0.05905584129316679
+   ]
+  },
+  "SC@7": {
+   "acc": 0.8717854518736223,
+   "equiv": 8047.607837419612,
+   "delta_vs_agree2": [
+    -0.068791329904482,
+    -0.08174136664217486,
+    -0.05648190668626012
+   ]
+  },
+  "SC@8": {
+   "acc": 0.8761021307861866,
+   "equiv": 9197.266099908129,
+   "delta_vs_agree2": [
+    -0.06447465099191771,
+    -0.07788850110213079,
+    -0.05143277002204261
+   ]
+  },
+  "AC(0.8)": {
+   "acc": 0.8714180749448934,
+   "equiv": 4412.0116364692885,
+   "delta_vs_agree2": [
+    -0.06915870683321088,
+    -0.08238886847905952,
+    -0.05639235855988244
+   ]
+  },
+  "AC(0.8)+think@2": {
+   "acc": 0.9405767817781043,
+   "equiv": 14776.582078233669,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@3": {
+   "acc": 0.9403930933137399,
+   "equiv": 15250.10020965791,
+   "delta_vs_agree2": [
+    -0.0001836884643644379,
+    -0.0005510653930933138,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@4": {
+   "acc": 0.9340558412931668,
+   "equiv": 12428.038051228672,
+   "delta_vs_agree2": [
+    -0.006520940484937546,
+    -0.009184423218221896,
+    -0.004132990448199853
+   ]
+  },
+  "AC(0.9)": {
+   "acc": 0.8757347538574578,
+   "equiv": 6165.396916034446,
+   "delta_vs_agree2": [
+    -0.06484202792064658,
+    -0.07834542615723733,
+    -0.05198383541513593
+   ]
+  },
+  "AC(0.9)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 34909.83373470687,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.9)+think@3": {
+   "acc": 0.9542615723732549,
+   "equiv": 18891.999423466008,
+   "delta_vs_agree2": [
+    0.013684790595150624,
+    0.010929463629684055,
+    0.016623806024981633
+   ]
+  },
+  "AC(0.9)+think@4": {
+   "acc": 0.9540778839088905,
+   "equiv": 19471.496539687407,
+   "delta_vs_agree2": [
+    0.013501102130786187,
+    0.010745775165319618,
+    0.0162587252020573
+   ]
+  },
+  "AC(0.95)": {
+   "acc": 0.8760102865540044,
+   "equiv": 7111.669323298678,
+   "delta_vs_agree2": [
+    -0.06456649522409992,
+    -0.07806989346069067,
+    -0.05152461425422483
+   ]
+  },
+  "AC(0.95)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 34909.83373470687,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.95)+think@3": {
+   "acc": 0.9603232916972814,
+   "equiv": 36059.49199719538,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.95)+think@4": {
+   "acc": 0.9572005878030859,
+   "equiv": 21701.026687494537,
+   "delta_vs_agree2": [
+    0.016623806024981633,
+    0.012950036737692874,
+    0.020575404114621592
+   ]
+  },
+  "AC(0.99)": {
+   "acc": 0.8761021307861866,
+   "equiv": 8341.71692481839,
+   "delta_vs_agree2": [
+    -0.06447465099191771,
+    -0.07788850110213079,
+    -0.05143277002204261
+   ]
+  },
+  "AC(0.99)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 34909.83373470687,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.99)+think@3": {
+   "acc": 0.9603232916972814,
+   "equiv": 36059.49199719538,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.99)+think@4": {
+   "acc": 0.9603232916972814,
+   "equiv": 37209.1502596839,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "always think": {
+   "acc": 0.9603232916972814,
+   "equiv": 32610.51720972983,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  }
+ },
+ "frontier": [
+  {
+   "R": 4000,
+   "acc": 0.7907788390889052,
+   "equiv": 1150.6473977249664
+  },
+  {
+   "R": 8000,
+   "acc": 0.7907788390889052,
+   "equiv": 1150.6473977249664
+  },
+  {
+   "R": 16000,
+   "acc": 0.8075863335782513,
+   "equiv": 1396.4326338052047
+  },
+  {
+   "R": 24000,
+   "acc": 0.8096987509184423,
+   "equiv": 1472.383530754564
+  },
+  {
+   "R": 32000,
+   "acc": 0.8292615723732549,
+   "equiv": 2379.59416181811
+  },
+  {
+   "R": 48000,
+   "acc": 0.863243938280676,
+   "equiv": 3899.246677291924
+  },
+  {
+   "R": 64000,
+   "acc": 0.8943791329904482,
+   "equiv": 7482.135276499471
+  },
+  {
+   "R": 96000,
+   "acc": 0.9239529757531227,
+   "equiv": 11234.58117973906
+  }
+ ],
+ "matched": [
+  {
+   "seed": 0,
+   "acc": 0.9448016164584864,
+   "equiv": 14828.426778364617,
+   "R": [
+    209675.72260947726,
+    368276.6581960034
+   ],
+   "delta_vs_agree2": [
+    0.004224834680382072,
+    0.0001836884643644379,
+    0.008357825128581926
+   ],
+   "actions": {
+    "stop@2": 1991,
+    "stop@3": 4195,
+    "stop@5": 1092,
+    "stop@6": 376,
+    "stop@4": 950,
+    "think:q32b_think@8": 624,
+    "think:q32b_think@7": 208,
+    "think:q32b_think@6": 220,
+    "stop@7": 263,
+    "think:q32b_think@3": 247,
+    "think:q32b_think@5": 366,
+    "stop@8": 218,
+    "think:q32b_think@4": 137,
+    "think:q32b_think@2": 1
+   }
+  },
+  {
+   "seed": 1,
+   "acc": 0.9444342395297576,
+   "equiv": 14563.816924343886,
+   "R": [
+    267815.15620037477,
+    295919.7331034693
+   ],
+   "delta_vs_agree2": [
+    0.003857457751653196,
+    -0.0004592211609110948,
+    0.008174136664217487
+   ],
+   "actions": {
+    "stop@8": 240,
+    "stop@7": 297,
+    "stop@6": 334,
+    "stop@5": 910,
+    "stop@3": 2091,
+    "think:q32b_think@8": 520,
+    "stop@2": 3447,
+    "stop@4": 1802,
+    "think:q32b_think@3": 133,
+    "think:q32b_think@4": 396,
+    "think:q32b_think@2": 5,
+    "think:q32b_think@5": 304,
+    "think:q32b_think@7": 179,
+    "think:q32b_think@6": 230
+   }
+  },
+  {
+   "seed": 2,
+   "acc": 0.9426891991182954,
+   "equiv": 14373.47393929234,
+   "R": [
+    250752.31351567752,
+    288676.0527428261
+   ],
+   "delta_vs_agree2": [
+    0.002112417340191036,
+    -0.002387950036737693,
+    0.006337252020573108
+   ],
+   "actions": {
+    "stop@5": 1013,
+    "stop@8": 229,
+    "stop@7": 232,
+    "stop@6": 331,
+    "stop@4": 1435,
+    "stop@3": 2894,
+    "stop@2": 2997,
+    "think:q32b_think@6": 391,
+    "think:q32b_think@3": 118,
+    "think:q32b_think@5": 261,
+    "think:q32b_think@8": 447,
+    "think:q32b_think@4": 397,
+    "think:q32b_think@7": 140,
+    "think:q32b_think@2": 3
+   }
+  },
+  {
+   "seed": 3,
+   "acc": 0.9433321087435709,
+   "equiv": 14789.907917129956,
+   "R": [
+    297679.4449697336,
+    308943.665263941
+   ],
+   "delta_vs_agree2": [
+    0.002755326965466569,
+    -0.002112417340191036,
+    0.007531227038941954
+   ],
+   "actions": {
+    "stop@2": 3282,
+    "stop@3": 2624,
+    "stop@8": 199,
+    "stop@7": 275,
+    "stop@6": 403,
+    "stop@4": 1575,
+    "stop@5": 756,
+    "think:q32b_think@7": 249,
+    "think:q32b_think@5": 359,
+    "think:q32b_think@6": 212,
+    "think:q32b_think@4": 290,
+    "think:q32b_think@8": 661,
+    "think:q32b_think@2": 3
+   }
+  }
+ ],
+ "matched_acc": [
+  {
+   "seed": 0,
+   "acc": 0.9356171932402645,
+   "equiv": 13166.091453002167,
+   "R": [
+    122453.90315791618,
+    215367.45691742294
+   ],
+   "delta_acc": [
+    -0.004959588537839824,
+    -0.009735488611315209,
+    -9.184423218221895e-05
+   ],
+   "rel_cost": [
+    -0.10898938717389867,
+    -0.14274393354401801,
+    -0.07372849351896534
+   ]
+  },
+  {
+   "seed": 1,
+   "acc": 0.9377296105804556,
+   "equiv": 12859.460458165413,
+   "R": [
+    128763.81386976562,
+    142514.1213197139
+   ],
+   "delta_acc": [
+    -0.002847171197648788,
+    -0.007531227038941954,
+    0.0018368846436443791
+   ],
+   "rel_cost": [
+    -0.12974053200653404,
+    -0.16560395904727204,
+    -0.09532283731871877
+   ]
+  },
+  {
+   "seed": 2,
+   "acc": 0.9364437913299045,
+   "equiv": 12809.857169532603,
+   "R": [
+    122551.04267303023,
+    142163.9738900473
+   ],
+   "delta_acc": [
+    -0.004132990448199853,
+    -0.009186719324026451,
+    0.0008265980896399706
+   ],
+   "rel_cost": [
+    -0.13309741713532697,
+    -0.1681467862102402,
+    -0.09717950555075175
+   ]
+  },
+  {
+   "seed": 3,
+   "acc": 0.9346069066862601,
+   "equiv": 12886.646277828833,
+   "R": [
+    171552.75460171435,
+    130081.31657895725
+   ],
+   "delta_acc": [
+    -0.005969875091844232,
+    -0.011572373254959589,
+    -0.0005510653930933138
+   ],
+   "rel_cost": [
+    -0.12790074121327172,
+    -0.16479919562013812,
+    -0.08909805666181518
+   ]
+  }
+ ]
+}

analysis/bayes_sc_2026-10-25/v8/sc32_gpu_s.log

@@ -0,0 +1,45 @@
+2026-10-25 14:12:15,490 online_rl_train INFO: ThinkBooster math evaluator loaded
+reference agree2 (rotation pairs, -> q32b_think): acc 0.941 equiv 14777
+cost model gpu_s: 1 equiv = 0.0006341 GPU-seconds; agree2 costs 9.37 GPU-seconds per problem
+  agree2                 acc 0.941 equiv  14777 (    +0; 9.37 GPU-seconds) | vs agree2 +0.0 [+0.0, +0.0]
+  SC@1                   acc 0.791 equiv   1150 (-13627; 0.729 GPU-seconds) | vs agree2 -15.0 [-16.3, -13.7]
+  SC@2                   acc 0.812 equiv   2299 (-12477; 1.458 GPU-seconds) | vs agree2 -12.9 [-14.2, -11.5]
+  SC@3                   acc 0.844 equiv   3449 (-11328; 2.187 GPU-seconds) | vs agree2 -9.6 [-10.9, -8.4]
+  SC@4                   acc 0.855 equiv   4599 (-10178; 2.916 GPU-seconds) | vs agree2 -8.5 [-9.8, -7.3]
+  SC@5                   acc 0.863 equiv   5748 ( -9028; 3.645 GPU-seconds) | vs agree2 -7.8 [-9.1, -6.5]
+  SC@6                   acc 0.869 equiv   6898 ( -7879; 4.374 GPU-seconds) | vs agree2 -7.1 [-8.4, -5.9]
+  SC@7                   acc 0.872 equiv   8048 ( -6729; 5.103 GPU-seconds) | vs agree2 -6.9 [-8.2, -5.6]
+  SC@8                   acc 0.876 equiv   9197 ( -5579; 5.832 GPU-seconds) | vs agree2 -6.4 [-7.8, -5.1]
+  AC(0.8)                acc 0.871 equiv   4412 (-10365; 2.798 GPU-seconds) | vs agree2 -6.9 [-8.2, -5.6]
+  AC(0.8)+think@2        acc 0.941 equiv  14777 (    +0; 9.37 GPU-seconds) | vs agree2 +0.0 [+0.0, +0.0]
+  AC(0.8)+think@3        acc 0.940 equiv  15250 (  +474; 9.67 GPU-seconds) | vs agree2 -0.0 [-0.1, +0.0]
+  AC(0.8)+think@4        acc 0.934 equiv  12428 ( -2349; 7.881 GPU-seconds) | vs agree2 -0.7 [-0.9, -0.4]
+  AC(0.9)                acc 0.876 equiv   6165 ( -8611; 3.91 GPU-seconds) | vs agree2 -6.5 [-7.8, -5.2]
+  AC(0.9)+think@2        acc 0.960 equiv  34910 (+20133; 22.14 GPU-seconds) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.9)+think@3        acc 0.954 equiv  18892 ( +4115; 11.98 GPU-seconds) | vs agree2 +1.4 [+1.1, +1.7]
+  AC(0.9)+think@4        acc 0.954 equiv  19471 ( +4695; 12.35 GPU-seconds) | vs agree2 +1.4 [+1.1, +1.6]
+  AC(0.95)               acc 0.876 equiv   7112 ( -7665; 4.51 GPU-seconds) | vs agree2 -6.5 [-7.8, -5.2]
+  AC(0.95)+think@2       acc 0.960 equiv  34910 (+20133; 22.14 GPU-seconds) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.95)+think@3       acc 0.960 equiv  36059 (+21283; 22.87 GPU-seconds) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.95)+think@4       acc 0.957 equiv  21701 ( +6924; 13.76 GPU-seconds) | vs agree2 +1.7 [+1.3, +2.1]
+  AC(0.99)               acc 0.876 equiv   8342 ( -6435; 5.29 GPU-seconds) | vs agree2 -6.4 [-7.8, -5.1]
+  AC(0.99)+think@2       acc 0.960 equiv  34910 (+20133; 22.14 GPU-seconds) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.99)+think@3       acc 0.960 equiv  36059 (+21283; 22.87 GPU-seconds) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.99)+think@4       acc 0.960 equiv  37209 (+22433; 23.59 GPU-seconds) | vs agree2 +2.0 [+1.3, +2.7]
+  always think           acc 0.960 equiv  32611 (+17834; 20.68 GPU-seconds) | vs agree2 +2.0 [+1.3, +2.7]
+R    4000: acc 0.791 equiv   1151
+R    8000: acc 0.791 equiv   1151
+R   16000: acc 0.808 equiv   1396
+R   24000: acc 0.810 equiv   1472
+R   32000: acc 0.829 equiv   2380
+R   48000: acc 0.863 equiv   3899
+R   64000: acc 0.894 equiv   7482
+R   96000: acc 0.924 equiv  11235
+parity: adaptive-SC Bayes at agree2's cost acc 0.945 equiv 14828 (+52) | vs agree2 +0.4 [+0.0, +0.8] | R [209676, 368277] | {'stop@3': 4195, 'stop@2': 1991, 'stop@5': 1092, 'stop@4': 950, 'think:q32b_think@8': 624, 'stop@6': 376, 'think:q32b_think@5': 366, 'stop@7': 263}
+split 1: adaptive-SC Bayes at agree2's cost acc 0.944 equiv 14564 (-213) | vs agree2 +0.4 [-0.0, +0.8] | R [267815, 295920] | {'stop@2': 3447, 'stop@3': 2091, 'stop@4': 1802, 'stop@5': 910, 'think:q32b_think@8': 520, 'think:q32b_think@4': 396, 'stop@6': 334, 'think:q32b_think@5': 304}
+split 2: adaptive-SC Bayes at agree2's cost acc 0.943 equiv 14373 (-403) | vs agree2 +0.2 [-0.2, +0.6] | R [250752, 288676] | {'stop@2': 2997, 'stop@3': 2894, 'stop@4': 1435, 'stop@5': 1013, 'think:q32b_think@8': 447, 'think:q32b_think@4': 397, 'think:q32b_think@6': 391, 'stop@6': 331}
+split 3: adaptive-SC Bayes at agree2's cost acc 0.943 equiv 14790 (+13) | vs agree2 +0.3 [-0.2, +0.8] | R [297679, 308944] | {'stop@2': 3282, 'stop@3': 2624, 'stop@4': 1575, 'stop@5': 756, 'think:q32b_think@8': 661, 'stop@6': 403, 'think:q32b_think@5': 359, 'think:q32b_think@4': 290}
+parity at agree2's accuracy: acc 0.936 cost 8.349 GPU-seconds vs 9.37 | cost -10.9% [-14.3, -7.4] | acc -0.5 [-1.0, -0.0]
+split 1 at agree2's accuracy: acc 0.938 cost 8.154 GPU-seconds vs 9.37 | cost -13.0% [-16.6, -9.5] | acc -0.3 [-0.8, +0.2]
+split 2 at agree2's accuracy: acc 0.936 cost 8.123 GPU-seconds vs 9.37 | cost -13.3% [-16.8, -9.7] | acc -0.4 [-0.9, +0.1]
+split 3 at agree2's accuracy: acc 0.935 cost 8.172 GPU-seconds vs 9.37 | cost -12.8% [-16.5, -8.9] | acc -0.6 [-1.2, -0.1]

analysis/bayes_sc_2026-10-25/v8/sc32_gpu_s.provenance.json

@@ -0,0 +1,90 @@
+{
+ "started_at": "2026-10-25T14:12:13+04:00",
+ "host": "ws006600",
+ "user": "vlad.smirnov",
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
+ "argv": [
+  "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
+  "--pool-dir",
+  "workdir/selector/pool_mth_k8_chat",
+  "--uhead",
+  "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+  "--task-prior",
+  "workdir/bayes/qprior/mth_qprior.json",
+  "--cost-model",
+  "gpu_s",
+  "--tier",
+  "q32b_think=workdir/bayes/q32b_think_tier.json@qwen3-32b-fp8:think",
+  "--match-acc",
+  "--out",
+  "workdir/bayes/v8/sc32_gpu_s.json"
+ ],
+ "git": {
+  "commit": "f935c73993aae0c1137fdec2866aef2005007331",
+  "branch": "HEAD",
+  "subject": "Per-problem arrays in the Bayesian controllers; paired bootstrap against the hull mixture in currency_table",
+  "dirty": false,
+  "dirty_files": []
+ },
+ "gpus": {
+  "visible": "",
+  "devices": [
+   {
+    "index": "0",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
+   },
+   {
+    "index": "1",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
+   }
+  ]
+ },
+ "packages": {
+  "python": "3.11.15",
+  "torch": "2.9.1",
+  "vllm": "0.15.1",
+  "transformers": "4.57.6",
+  "speculators": "0.4.0.dev49",
+  "peft": null,
+  "accelerate": "1.13.0",
+  "numpy": "2.2.6",
+  "huggingface_hub": "0.36.2",
+  "openai": "2.32.0",
+  "wandb": "0.25.1"
+ },
+ "env": {
+  "CUDA_VISIBLE_DEVICES": ""
+ },
+ "models": [],
+ "files": {
+  "pool": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
+   "bytes": 65136887,
+   "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
+  },
+  "uhead": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+   "bytes": 9369632,
+   "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
+  },
+  "task_prior": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/qprior/mth_qprior.json",
+   "bytes": 28987,
+   "sha256": "897f405c643430dbe97023d9f8985e108330b82ed412a25ab2a4b84c992fbbd6"
+  },
+  "tier:q32b_think": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/q32b_think_tier.json",
+   "bytes": 26525,
+   "sha256": "4eef554b7fa4edff6cf11b63aadb44267d211bc25285eb260dcc1b45d2cf3a1f"
+  }
+ },
+ "finished_at": "2026-10-25T14:13:20+04:00",
+ "wall_secs": 66.8,
+ "n_problems": 1361
+}

analysis/bayes_sc_2026-10-25/v8/sc32_tokens.arrays.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa60b51d4c54c38e2706cf4416427395a1a13fd4a199ed9f8b235e4042be2527
+size 183873

analysis/bayes_sc_2026-10-25/v8/sc32_tokens.json

@@ -0,0 +1,500 @@
+{
+ "ref": {
+  "acc": 0.9405767817781043,
+  "equiv": 4283.8179463629685
+ },
+ "cost_model": {
+  "currency": "tokens",
+  "unit": "base tokens",
+  "per_equiv": 1.0,
+  "scenario": null,
+  "base": "qwen3-8b:nothink"
+ },
+ "baselines": {
+  "agree2": {
+   "acc": 0.9405767817781043,
+   "equiv": 4283.8179463629685,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "SC@1": {
+   "acc": 0.7907788390889052,
+   "equiv": 1124.9408523144746,
+   "delta_vs_agree2": [
+    -0.14979794268919913,
+    -0.16256429096252756,
+    -0.1369374540778839
+   ]
+  },
+  "SC@2": {
+   "acc": 0.8118111682586333,
+   "equiv": 2249.881704628949,
+   "delta_vs_agree2": [
+    -0.12876561351947097,
+    -0.14226671565025717,
+    -0.11544819985304923
+   ]
+  },
+  "SC@3": {
+   "acc": 0.8444158706833211,
+   "equiv": 3374.822556943424,
+   "delta_vs_agree2": [
+    -0.09616091109478325,
+    -0.10911324393828067,
+    -0.08357825128581925
+   ]
+  },
+  "SC@4": {
+   "acc": 0.8553453343130052,
+   "equiv": 4499.763409257898,
+   "delta_vs_agree2": [
+    -0.0852314474650992,
+    -0.09809652828802351,
+    -0.07274063188831742
+   ]
+  },
+  "SC@5": {
+   "acc": 0.8630602498163116,
+   "equiv": 5624.704261572373,
+   "delta_vs_agree2": [
+    -0.0775165319617928,
+    -0.09056070903747244,
+    -0.06493387215282881
+   ]
+  },
+  "SC@6": {
+   "acc": 0.8693056576047025,
+   "equiv": 6749.645113886848,
+   "delta_vs_agree2": [
+    -0.07127112417340191,
+    -0.08440714548126378,
+    -0.05905584129316679
+   ]
+  },
+  "SC@7": {
+   "acc": 0.8717854518736223,
+   "equiv": 7874.585966201323,
+   "delta_vs_agree2": [
+    -0.068791329904482,
+    -0.08174136664217486,
+    -0.05648190668626012
+   ]
+  },
+  "SC@8": {
+   "acc": 0.8761021307861866,
+   "equiv": 8999.526818515797,
+   "delta_vs_agree2": [
+    -0.06447465099191771,
+    -0.07788850110213079,
+    -0.05143277002204261
+   ]
+  },
+  "AC(0.8)": {
+   "acc": 0.8714180749448934,
+   "equiv": 4328.9513225569435,
+   "delta_vs_agree2": [
+    -0.06915870683321088,
+    -0.08238886847905952,
+    -0.05639235855988244
+   ]
+  },
+  "AC(0.8)+think@2": {
+   "acc": 0.9405767817781043,
+   "equiv": 4283.8179463629685,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@3": {
+   "acc": 0.9403930933137399,
+   "equiv": 4758.662803085966,
+   "delta_vs_agree2": [
+    -0.0001836884643644379,
+    -0.0005510653930933138,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@4": {
+   "acc": 0.9340558412931668,
+   "equiv": 4695.218185157972,
+   "delta_vs_agree2": [
+    -0.006520940484937546,
+    -0.009184423218221896,
+    -0.004132990448199853
+   ]
+  },
+  "AC(0.9)": {
+   "acc": 0.8757347538574578,
+   "equiv": 6046.053545187362,
+   "delta_vs_agree2": [
+    -0.06484202792064658,
+    -0.07834542615723733,
+    -0.05198383541513593
+   ]
+  },
+  "AC(0.9)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 7567.180749448935,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.9)+think@3": {
+   "acc": 0.9542615723732549,
+   "equiv": 5892.238997060985,
+   "delta_vs_agree2": [
+    0.013684790595150624,
+    0.010929463629684055,
+    0.016623806024981633
+   ]
+  },
+  "AC(0.9)+think@4": {
+   "acc": 0.9540778839088905,
+   "equiv": 6471.090558412931,
+   "delta_vs_agree2": [
+    0.013501102130786187,
+    0.010745775165319618,
+    0.0162587252020573
+   ]
+  },
+  "AC(0.95)": {
+   "acc": 0.8760102865540044,
+   "equiv": 6970.480804555474,
+   "delta_vs_agree2": [
+    -0.06456649522409992,
+    -0.07806989346069067,
+    -0.05152461425422483
+   ]
+  },
+  "AC(0.95)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 7567.180749448935,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.95)+think@3": {
+   "acc": 0.9603232916972814,
+   "equiv": 8692.121601763409,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.95)+think@4": {
+   "acc": 0.9572005878030859,
+   "equiv": 7287.702479794269,
+   "delta_vs_agree2": [
+    0.016623806024981633,
+    0.012950036737692874,
+    0.020575404114621592
+   ]
+  },
+  "AC(0.99)": {
+   "acc": 0.8761021307861866,
+   "equiv": 8168.442964731815,
+   "delta_vs_agree2": [
+    -0.06447465099191771,
+    -0.07788850110213079,
+    -0.05143277002204261
+   ]
+  },
+  "AC(0.99)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 7567.180749448935,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.99)+think@3": {
+   "acc": 0.9603232916972814,
+   "equiv": 8692.121601763409,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.99)+think@4": {
+   "acc": 0.9603232916972814,
+   "equiv": 9817.062454077884,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "always think": {
+   "acc": 0.9603232916972814,
+   "equiv": 5317.299044819985,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  }
+ },
+ "frontier": [
+  {
+   "R": 4000,
+   "acc": 0.7909625275532697,
+   "equiv": 1126.719911829537
+  },
+  {
+   "R": 8000,
+   "acc": 0.7910543717854519,
+   "equiv": 1127.2466936076414
+  },
+  {
+   "R": 16000,
+   "acc": 0.93708670095518,
+   "equiv": 3569.6509919177074
+  },
+  {
+   "R": 24000,
+   "acc": 0.9426891991182954,
+   "equiv": 3761.4378398236595
+  },
+  {
+   "R": 32000,
+   "acc": 0.9488427626745041,
+   "equiv": 3929.1393093313736
+  },
+  {
+   "R": 48000,
+   "acc": 0.9541697281410727,
+   "equiv": 4187.878085966201
+  },
+  {
+   "R": 64000,
+   "acc": 0.9560066127847171,
+   "equiv": 4303.237656135195
+  },
+  {
+   "R": 96000,
+   "acc": 0.9575679647318148,
+   "equiv": 4473.639070536371
+  }
+ ],
+ "matched": [
+  {
+   "seed": 0,
+   "acc": 0.9546289493019838,
+   "equiv": 4215.773769287289,
+   "R": [
+    91824.58538325278,
+    40573.87526947939
+   ],
+   "delta_vs_agree2": [
+    0.0140521675238795,
+    0.008541513592946363,
+    0.019930198383541514
+   ],
+   "actions": {
+    "stop@2": 4261,
+    "stop@5": 253,
+    "think:q32b_think@2": 144,
+    "stop@3": 144,
+    "think:q32b_think@0": 3576,
+    "stop@6": 40,
+    "think:q32b_think@8": 17,
+    "think:q32b_think@7": 20,
+    "think:q32b_think@5": 9,
+    "stop@8": 28,
+    "stop@7": 48,
+    "think:q32b_think@3": 46,
+    "think:q32b_think@6": 25,
+    "think:q32b_think@4": 21,
+    "think:q32b_think@1": 522,
+    "stop@1": 1734
+   }
+  },
+  {
+   "seed": 1,
+   "acc": 0.9538941954445261,
+   "equiv": 4314.635947832477,
+   "R": [
+    67328.52078761288,
+    55381.77637001384
+   ],
+   "delta_vs_agree2": [
+    0.01331741366642175,
+    0.007896307861866275,
+    0.019103600293901544
+   ],
+   "actions": {
+    "stop@8": 62,
+    "think:q32b_think@3": 104,
+    "think:q32b_think@2": 318,
+    "stop@3": 583,
+    "stop@2": 5062,
+    "think:q32b_think@0": 3688,
+    "stop@5": 245,
+    "stop@7": 57,
+    "stop@4": 39,
+    "think:q32b_think@4": 26,
+    "stop@6": 53,
+    "think:q32b_think@7": 38,
+    "think:q32b_think@8": 31,
+    "think:q32b_think@5": 6,
+    "think:q32b_think@6": 15,
+    "stop@1": 561
+   }
+  },
+  {
+   "seed": 2,
+   "acc": 0.9540778839088905,
+   "equiv": 4266.510562086701,
+   "R": [
+    79283.1784206788,
+    44078.78098259121
+   ],
+   "delta_vs_agree2": [
+    0.013501102130786187,
+    0.008082292432035268,
+    0.019289584864070525
+   ],
+   "actions": {
+    "think:q32b_think@2": 388,
+    "think:q32b_think@3": 227,
+    "stop@7": 28,
+    "stop@6": 39,
+    "stop@5": 141,
+    "stop@4": 76,
+    "stop@3": 173,
+    "stop@2": 5479,
+    "think:q32b_think@4": 24,
+    "think:q32b_think@0": 3600,
+    "think:q32b_think@7": 4,
+    "think:q32b_think@8": 7,
+    "stop@8": 57,
+    "think:q32b_think@5": 10,
+    "think:q32b_think@6": 14,
+    "stop@1": 621
+   }
+  },
+  {
+   "seed": 3,
+   "acc": 0.9544452608376194,
+   "equiv": 4242.00922116091,
+   "R": [
+    49402.412840278674,
+    66141.97703218769
+   ],
+   "delta_vs_agree2": [
+    0.013868479059515063,
+    0.008541513592946363,
+    0.019562821454812637
+   ],
+   "actions": {
+    "stop@2": 5320,
+    "think:q32b_think@1": 399,
+    "stop@1": 616,
+    "stop@5": 148,
+    "think:q32b_think@0": 3648,
+    "think:q32b_think@3": 53,
+    "stop@6": 21,
+    "think:q32b_think@4": 26,
+    "stop@7": 34,
+    "think:q32b_think@8": 16,
+    "think:q32b_think@2": 335,
+    "think:q32b_think@5": 4,
+    "stop@8": 21,
+    "stop@4": 29,
+    "think:q32b_think@6": 3,
+    "think:q32b_think@7": 6,
+    "stop@3": 209
+   }
+  }
+ ],
+ "matched_acc": [
+  {
+   "seed": 0,
+   "acc": 0.9426891991182954,
+   "equiv": 3761.095481263777,
+   "R": [
+    22358.663294702732,
+    22482.003544281713
+   ],
+   "delta_acc": [
+    0.002112417340191036,
+    -0.00395159808963997,
+    0.009000734753857457
+   ],
+   "rel_cost": [
+    -0.12202256763572124,
+    -0.14920852463555445,
+    -0.09306146425335104
+   ]
+  },
+  {
+   "seed": 1,
+   "acc": 0.942138133725202,
+   "equiv": 3769.3541513592945,
+   "R": [
+    23367.790099632886,
+    27279.94519816034
+   ],
+   "delta_acc": [
+    0.0015613519470977222,
+    -0.004410819250551065,
+    0.00789860396767083
+   ],
+   "rel_cost": [
+    -0.12009469156840848,
+    -0.14815632663270598,
+    -0.09143833511078628
+   ]
+  },
+  {
+   "seed": 2,
+   "acc": 0.9412196914033799,
+   "equiv": 3721.1166605437184,
+   "R": [
+    20256.816931905476,
+    21697.90831770529
+   ],
+   "delta_acc": [
+    0.0006429096252755327,
+    -0.005418809698750918,
+    0.007531227038941954
+   ],
+   "rel_cost": [
+    -0.13135508858330291,
+    -0.16029160822478866,
+    -0.10237103313659894
+   ]
+  },
+  {
+   "seed": 3,
+   "acc": 0.9424136664217487,
+   "equiv": 3732.863409257898,
+   "R": [
+    26373.01397547303,
+    22770.81386127522
+   ],
+   "delta_acc": [
+    0.0018368846436443791,
+    -0.004227130786186627,
+    0.008451965466568687
+   ],
+   "rel_cost": [
+    -0.12861296721837565,
+    -0.15635279023593773,
+    -0.10022076741312545
+   ]
+  }
+ ]
+}

analysis/bayes_sc_2026-10-25/v8/sc32_tokens.log

@@ -0,0 +1,45 @@
+2026-10-25 14:12:15,329 online_rl_train INFO: ThinkBooster math evaluator loaded
+reference agree2 (rotation pairs, -> q32b_think): acc 0.941 equiv 4284
+cost model tokens: 1 equiv = 1 base tokens; agree2 costs 4284 base tokens per problem
+  agree2                 acc 0.941 equiv   4284 (    +0; 4284 base tokens) | vs agree2 +0.0 [+0.0, +0.0]
+  SC@1                   acc 0.791 equiv   1125 ( -3159; 1125 base tokens) | vs agree2 -15.0 [-16.3, -13.7]
+  SC@2                   acc 0.812 equiv   2250 ( -2034; 2250 base tokens) | vs agree2 -12.9 [-14.2, -11.5]
+  SC@3                   acc 0.844 equiv   3375 (  -909; 3375 base tokens) | vs agree2 -9.6 [-10.9, -8.4]
+  SC@4                   acc 0.855 equiv   4500 (  +216; 4500 base tokens) | vs agree2 -8.5 [-9.8, -7.3]
+  SC@5                   acc 0.863 equiv   5625 ( +1341; 5625 base tokens) | vs agree2 -7.8 [-9.1, -6.5]
+  SC@6                   acc 0.869 equiv   6750 ( +2466; 6750 base tokens) | vs agree2 -7.1 [-8.4, -5.9]
+  SC@7                   acc 0.872 equiv   7875 ( +3591; 7875 base tokens) | vs agree2 -6.9 [-8.2, -5.6]
+  SC@8                   acc 0.876 equiv   9000 ( +4716; 9000 base tokens) | vs agree2 -6.4 [-7.8, -5.1]
+  AC(0.8)                acc 0.871 equiv   4329 (   +45; 4329 base tokens) | vs agree2 -6.9 [-8.2, -5.6]
+  AC(0.8)+think@2        acc 0.941 equiv   4284 (    +0; 4284 base tokens) | vs agree2 +0.0 [+0.0, +0.0]
+  AC(0.8)+think@3        acc 0.940 equiv   4759 (  +475; 4759 base tokens) | vs agree2 -0.0 [-0.1, +0.0]
+  AC(0.8)+think@4        acc 0.934 equiv   4695 (  +411; 4695 base tokens) | vs agree2 -0.7 [-0.9, -0.4]
+  AC(0.9)                acc 0.876 equiv   6046 ( +1762; 6046 base tokens) | vs agree2 -6.5 [-7.8, -5.2]
+  AC(0.9)+think@2        acc 0.960 equiv   7567 ( +3283; 7567 base tokens) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.9)+think@3        acc 0.954 equiv   5892 ( +1608; 5892 base tokens) | vs agree2 +1.4 [+1.1, +1.7]
+  AC(0.9)+think@4        acc 0.954 equiv   6471 ( +2187; 6471 base tokens) | vs agree2 +1.4 [+1.1, +1.6]
+  AC(0.95)               acc 0.876 equiv   6970 ( +2687; 6970 base tokens) | vs agree2 -6.5 [-7.8, -5.2]
+  AC(0.95)+think@2       acc 0.960 equiv   7567 ( +3283; 7567 base tokens) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.95)+think@3       acc 0.960 equiv   8692 ( +4408; 8692 base tokens) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.95)+think@4       acc 0.957 equiv   7288 ( +3004; 7288 base tokens) | vs agree2 +1.7 [+1.3, +2.1]
+  AC(0.99)               acc 0.876 equiv   8168 ( +3885; 8168 base tokens) | vs agree2 -6.4 [-7.8, -5.1]
+  AC(0.99)+think@2       acc 0.960 equiv   7567 ( +3283; 7567 base tokens) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.99)+think@3       acc 0.960 equiv   8692 ( +4408; 8692 base tokens) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.99)+think@4       acc 0.960 equiv   9817 ( +5533; 9817 base tokens) | vs agree2 +2.0 [+1.3, +2.7]
+  always think           acc 0.960 equiv   5317 ( +1033; 5317 base tokens) | vs agree2 +2.0 [+1.3, +2.7]
+R    4000: acc 0.791 equiv   1127
+R    8000: acc 0.791 equiv   1127
+R   16000: acc 0.937 equiv   3570
+R   24000: acc 0.943 equiv   3761
+R   32000: acc 0.949 equiv   3929
+R   48000: acc 0.954 equiv   4188
+R   64000: acc 0.956 equiv   4303
+R   96000: acc 0.958 equiv   4474
+parity: adaptive-SC Bayes at agree2's cost acc 0.955 equiv 4216 (-68) | vs agree2 +1.4 [+0.9, +2.0] | R [91825, 40574] | {'stop@2': 4261, 'think:q32b_think@0': 3576, 'stop@1': 1734, 'think:q32b_think@1': 522, 'stop@5': 253, 'think:q32b_think@2': 144, 'stop@3': 144, 'stop@7': 48}
+split 1: adaptive-SC Bayes at agree2's cost acc 0.954 equiv 4315 (+31) | vs agree2 +1.3 [+0.8, +1.9] | R [67329, 55382] | {'stop@2': 5062, 'think:q32b_think@0': 3688, 'stop@3': 583, 'stop@1': 561, 'think:q32b_think@2': 318, 'stop@5': 245, 'think:q32b_think@3': 104, 'stop@8': 62}
+split 2: adaptive-SC Bayes at agree2's cost acc 0.954 equiv 4267 (-17) | vs agree2 +1.4 [+0.8, +1.9] | R [79283, 44079] | {'stop@2': 5479, 'think:q32b_think@0': 3600, 'stop@1': 621, 'think:q32b_think@2': 388, 'think:q32b_think@3': 227, 'stop@3': 173, 'stop@5': 141, 'stop@4': 76}
+split 3: adaptive-SC Bayes at agree2's cost acc 0.954 equiv 4242 (-42) | vs agree2 +1.4 [+0.9, +2.0] | R [49402, 66142] | {'stop@2': 5320, 'think:q32b_think@0': 3648, 'stop@1': 616, 'think:q32b_think@1': 399, 'think:q32b_think@2': 335, 'stop@3': 209, 'stop@5': 148, 'think:q32b_think@3': 53}
+parity at agree2's accuracy: acc 0.943 cost 3761 base tokens vs 4284 | cost -12.2% [-14.9, -9.3] | acc +0.2 [-0.4, +0.9]
+split 1 at agree2's accuracy: acc 0.942 cost 3769 base tokens vs 4284 | cost -12.0% [-14.8, -9.1] | acc +0.2 [-0.4, +0.8]
+split 2 at agree2's accuracy: acc 0.941 cost 3721 base tokens vs 4284 | cost -13.1% [-16.0, -10.2] | acc +0.1 [-0.5, +0.8]
+split 3 at agree2's accuracy: acc 0.942 cost 3733 base tokens vs 4284 | cost -12.9% [-15.6, -10.0] | acc +0.2 [-0.4, +0.8]

analysis/bayes_sc_2026-10-25/v8/sc32_tokens.provenance.json

@@ -0,0 +1,90 @@
+{
+ "started_at": "2026-10-25T14:12:13+04:00",
+ "host": "ws006600",
+ "user": "vlad.smirnov",
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
+ "argv": [
+  "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
+  "--pool-dir",
+  "workdir/selector/pool_mth_k8_chat",
+  "--uhead",
+  "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+  "--task-prior",
+  "workdir/bayes/qprior/mth_qprior.json",
+  "--cost-model",
+  "tokens",
+  "--tier",
+  "q32b_think=workdir/bayes/q32b_think_tier.json",
+  "--match-acc",
+  "--out",
+  "workdir/bayes/v8/sc32_tokens.json"
+ ],
+ "git": {
+  "commit": "f935c73993aae0c1137fdec2866aef2005007331",
+  "branch": "HEAD",
+  "subject": "Per-problem arrays in the Bayesian controllers; paired bootstrap against the hull mixture in currency_table",
+  "dirty": false,
+  "dirty_files": []
+ },
+ "gpus": {
+  "visible": "",
+  "devices": [
+   {
+    "index": "0",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
+   },
+   {
+    "index": "1",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
+   }
+  ]
+ },
+ "packages": {
+  "python": "3.11.15",
+  "torch": "2.9.1",
+  "vllm": "0.15.1",
+  "transformers": "4.57.6",
+  "speculators": "0.4.0.dev49",
+  "peft": null,
+  "accelerate": "1.13.0",
+  "numpy": "2.2.6",
+  "huggingface_hub": "0.36.2",
+  "openai": "2.32.0",
+  "wandb": "0.25.1"
+ },
+ "env": {
+  "CUDA_VISIBLE_DEVICES": ""
+ },
+ "models": [],
+ "files": {
+  "pool": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
+   "bytes": 65136887,
+   "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
+  },
+  "uhead": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+   "bytes": 9369632,
+   "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
+  },
+  "task_prior": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/qprior/mth_qprior.json",
+   "bytes": 28987,
+   "sha256": "897f405c643430dbe97023d9f8985e108330b82ed412a25ab2a4b84c992fbbd6"
+  },
+  "tier:q32b_think": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/q32b_think_tier.json",
+   "bytes": 26525,
+   "sha256": "4eef554b7fa4edff6cf11b63aadb44267d211bc25285eb260dcc1b45d2cf3a1f"
+  }
+ },
+ "finished_at": "2026-10-25T14:12:56+04:00",
+ "wall_secs": 43.7,
+ "n_problems": 1361
+}

analysis/bayes_sc_2026-10-25/v8/sc32_usd_cheapest.arrays.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ba1d74f9e342a7f1d4e7eb678bdc03ee287fddf2c81be01eb242d1e834f7860
+size 189757

analysis/bayes_sc_2026-10-25/v8/sc32_usd_cheapest.json

@@ -0,0 +1,499 @@
+{
+ "ref": {
+  "acc": 0.9405767817781043,
+  "equiv": 5150.545609845701
+ },
+ "cost_model": {
+  "currency": "usd",
+  "unit": "USD",
+  "per_equiv": 2.0000000000000002e-07,
+  "scenario": "cheapest",
+  "base": "qwen3-8b:nothink"
+ },
+ "baselines": {
+  "agree2": {
+   "acc": 0.9405767817781043,
+   "equiv": 5150.545609845701,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "SC@1": {
+   "acc": 0.7907788390889052,
+   "equiv": 1149.6065393093313,
+   "delta_vs_agree2": [
+    -0.14979794268919913,
+    -0.16256429096252756,
+    -0.1369374540778839
+   ]
+  },
+  "SC@2": {
+   "acc": 0.8118111682586333,
+   "equiv": 2299.2130786186626,
+   "delta_vs_agree2": [
+    -0.12876561351947097,
+    -0.14226671565025717,
+    -0.11544819985304923
+   ]
+  },
+  "SC@3": {
+   "acc": 0.8444158706833211,
+   "equiv": 3448.819617927994,
+   "delta_vs_agree2": [
+    -0.09616091109478325,
+    -0.10911324393828067,
+    -0.08357825128581925
+   ]
+  },
+  "SC@4": {
+   "acc": 0.8553453343130052,
+   "equiv": 4598.426157237325,
+   "delta_vs_agree2": [
+    -0.0852314474650992,
+    -0.09809652828802351,
+    -0.07274063188831742
+   ]
+  },
+  "SC@5": {
+   "acc": 0.8630602498163116,
+   "equiv": 5748.032696546657,
+   "delta_vs_agree2": [
+    -0.0775165319617928,
+    -0.09056070903747244,
+    -0.06493387215282881
+   ]
+  },
+  "SC@6": {
+   "acc": 0.8693056576047025,
+   "equiv": 6897.639235855988,
+   "delta_vs_agree2": [
+    -0.07127112417340191,
+    -0.08440714548126378,
+    -0.05905584129316679
+   ]
+  },
+  "SC@7": {
+   "acc": 0.8717854518736223,
+   "equiv": 8047.245775165319,
+   "delta_vs_agree2": [
+    -0.068791329904482,
+    -0.08174136664217486,
+    -0.05648190668626012
+   ]
+  },
+  "SC@8": {
+   "acc": 0.8761021307861866,
+   "equiv": 9196.85231447465,
+   "delta_vs_agree2": [
+    -0.06447465099191771,
+    -0.07788850110213079,
+    -0.05143277002204261
+   ]
+  },
+  "AC(0.8)": {
+   "acc": 0.8714180749448934,
+   "equiv": 4411.8378260470245,
+   "delta_vs_agree2": [
+    -0.06915870683321088,
+    -0.08238886847905952,
+    -0.05639235855988244
+   ]
+  },
+  "AC(0.8)+think@2": {
+   "acc": 0.9405767817781043,
+   "equiv": 5150.545609845701,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@3": {
+   "acc": 0.9403930933137399,
+   "equiv": 5632.626047024248,
+   "delta_vs_agree2": [
+    -0.0001836884643644379,
+    -0.0005510653930933138,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@4": {
+   "acc": 0.9340558412931668,
+   "equiv": 5360.495435341661,
+   "delta_vs_agree2": [
+    -0.006520940484937546,
+    -0.009184423218221896,
+    -0.004132990448199853
+   ]
+  },
+  "AC(0.9)": {
+   "acc": 0.8757347538574578,
+   "equiv": 6165.147180382072,
+   "delta_vs_agree2": [
+    -0.06484202792064658,
+    -0.07834542615723733,
+    -0.05198383541513593
+   ]
+  },
+  "AC(0.9)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 9755.271271124173,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.9)+think@3": {
+   "acc": 0.9542615723732549,
+   "equiv": 6977.969140337987,
+   "delta_vs_agree2": [
+    0.013684790595150624,
+    0.010929463629684055,
+    0.016623806024981633
+   ]
+  },
+  "AC(0.9)+think@4": {
+   "acc": 0.9540778839088905,
+   "equiv": 7566.024435157972,
+   "delta_vs_agree2": [
+    0.013501102130786187,
+    0.010745775165319618,
+    0.0162587252020573
+   ]
+  },
+  "AC(0.95)": {
+   "acc": 0.8760102865540044,
+   "equiv": 7111.373874908156,
+   "delta_vs_agree2": [
+    -0.06456649522409992,
+    -0.07806989346069067,
+    -0.05152461425422483
+   ]
+  },
+  "AC(0.95)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 9755.271271124173,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.95)+think@3": {
+   "acc": 0.9603232916972814,
+   "equiv": 10904.877810433503,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.95)+think@4": {
+   "acc": 0.9572005878030859,
+   "equiv": 8506.850459221161,
+   "delta_vs_agree2": [
+    0.016623806024981633,
+    0.012950036737692874,
+    0.020575404114621592
+   ]
+  },
+  "AC(0.99)": {
+   "acc": 0.8761021307861866,
+   "equiv": 8341.35433504776,
+   "delta_vs_agree2": [
+    -0.06447465099191771,
+    -0.07788850110213079,
+    -0.05143277002204261
+   ]
+  },
+  "AC(0.99)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 9755.271271124173,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.99)+think@3": {
+   "acc": 0.9603232916972814,
+   "equiv": 10904.877810433503,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.99)+think@4": {
+   "acc": 0.9603232916972814,
+   "equiv": 12054.484349742836,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "always think": {
+   "acc": 0.9603232916972814,
+   "equiv": 7456.05819250551,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  }
+ },
+ "frontier": [
+  {
+   "R": 4000,
+   "acc": 0.7907788390889052,
+   "equiv": 1150.5956557678178
+  },
+  {
+   "R": 8000,
+   "acc": 0.7907788390889052,
+   "equiv": 1150.5956557678178
+  },
+  {
+   "R": 16000,
+   "acc": 0.8692138133725202,
+   "equiv": 3077.968547942689
+  },
+  {
+   "R": 24000,
+   "acc": 0.9411278471711977,
+   "equiv": 4937.058826230713
+  },
+  {
+   "R": 32000,
+   "acc": 0.9473732549595886,
+   "equiv": 5096.705244305658
+  },
+  {
+   "R": 48000,
+   "acc": 0.9528839088905217,
+   "equiv": 5359.611760653932
+  },
+  {
+   "R": 64000,
+   "acc": 0.955088170462895,
+   "equiv": 5517.903719691403
+  },
+  {
+   "R": 96000,
+   "acc": 0.9560984570168993,
+   "equiv": 5646.251644011756
+  }
+ ],
+ "matched": [
+  {
+   "seed": 0,
+   "acc": 0.9493019838354151,
+   "equiv": 5179.889915503306,
+   "R": [
+    44001.89742887743,
+    30172.35271204413
+   ],
+   "delta_vs_agree2": [
+    0.008725202057310802,
+    0.0029390154298310064,
+    0.014972905951506233
+   ],
+   "actions": {
+    "stop@1": 3414,
+    "stop@2": 2949,
+    "stop@4": 139,
+    "think:q32b_think@2": 271,
+    "think:q32b_think@0": 3576,
+    "stop@5": 160,
+    "stop@6": 39,
+    "stop@7": 65,
+    "think:q32b_think@7": 19,
+    "think:q32b_think@5": 12,
+    "think:q32b_think@3": 15,
+    "stop@8": 27,
+    "think:q32b_think@4": 31,
+    "think:q32b_think@6": 7,
+    "think:q32b_think@8": 3,
+    "stop@3": 161
+   }
+  },
+  {
+   "seed": 1,
+   "acc": 0.9442505510653931,
+   "equiv": 5077.934055841293,
+   "R": [
+    36742.20417604087,
+    30332.785879190633
+   ],
+   "delta_vs_agree2": [
+    0.0036737692872887582,
+    -0.002387950036737693,
+    0.010105161645848628
+   ],
+   "actions": {
+    "stop@2": 2279,
+    "think:q32b_think@3": 82,
+    "stop@4": 205,
+    "think:q32b_think@0": 3688,
+    "stop@1": 4215,
+    "stop@7": 73,
+    "think:q32b_think@2": 104,
+    "stop@6": 34,
+    "stop@3": 49,
+    "think:q32b_think@7": 11,
+    "think:q32b_think@5": 20,
+    "stop@5": 60,
+    "think:q32b_think@8": 14,
+    "think:q32b_think@4": 15,
+    "stop@8": 37,
+    "think:q32b_think@6": 2
+   }
+  },
+  {
+   "seed": 2,
+   "acc": 0.946730345334313,
+   "equiv": 5108.292932586333,
+   "R": [
+    43714.70455047698,
+    32839.90794532481
+   ],
+   "delta_vs_agree2": [
+    0.00615356355620867,
+    0.0002755326965466569,
+    0.012398971344599559
+   ],
+   "actions": {
+    "stop@5": 206,
+    "stop@6": 31,
+    "stop@2": 2611,
+    "stop@1": 3926,
+    "think:q32b_think@3": 73,
+    "think:q32b_think@2": 152,
+    "think:q32b_think@4": 18,
+    "think:q32b_think@0": 3600,
+    "stop@3": 119,
+    "think:q32b_think@8": 14,
+    "think:q32b_think@7": 11,
+    "stop@7": 46,
+    "think:q32b_think@5": 44,
+    "stop@8": 27,
+    "stop@4": 9,
+    "think:q32b_think@6": 1
+   }
+  },
+  {
+   "seed": 3,
+   "acc": 0.9483835415135929,
+   "equiv": 5101.276542983101,
+   "R": [
+    39017.86102873052,
+    32700.663999164928
+   ],
+   "delta_vs_agree2": [
+    0.007806759735488611,
+    0.002204261572373255,
+    0.013868479059515063
+   ],
+   "actions": {
+    "stop@1": 3963,
+    "stop@2": 2498,
+    "think:q32b_think@2": 258,
+    "stop@5": 99,
+    "stop@6": 22,
+    "think:q32b_think@0": 3648,
+    "stop@4": 169,
+    "stop@3": 77,
+    "stop@8": 24,
+    "stop@7": 40,
+    "think:q32b_think@5": 14,
+    "think:q32b_think@4": 14,
+    "think:q32b_think@8": 6,
+    "think:q32b_think@6": 5,
+    "think:q32b_think@7": 3,
+    "think:q32b_think@3": 48
+   }
+  }
+ ],
+ "matched_acc": [
+  {
+   "seed": 0,
+   "acc": 0.9410360029390155,
+   "equiv": 4936.350133174136,
+   "R": [
+    23403.76818773052,
+    23382.30415765745
+   ],
+   "delta_acc": [
+    0.0004592211609110948,
+    -0.005696638501102131,
+    0.007439382806759736
+   ],
+   "rel_cost": [
+    -0.041586948819968196,
+    -0.07185071633836701,
+    -0.00932253431323091
+   ]
+  },
+  {
+   "seed": 1,
+   "acc": 0.9412196914033799,
+   "equiv": 4956.526010286553,
+   "R": [
+    24460.01596270518,
+    28000.174242016714
+   ],
+   "delta_acc": [
+    0.0006429096252755327,
+    -0.005602498163115356,
+    0.007074301983835403
+   ],
+   "rel_cost": [
+    -0.03766971778451256,
+    -0.0694605956345558,
+    -0.0055554942161507
+   ]
+  },
+  {
+   "seed": 2,
+   "acc": 0.9386480529022777,
+   "equiv": 4866.483096069066,
+   "R": [
+    23805.320187558315,
+    22665.287727281033
+   ],
+   "delta_acc": [
+    -0.001928728875826598,
+    -0.008451965466568699,
+    0.004961884643644367
+   ],
+   "rel_cost": [
+    -0.055151926668434006,
+    -0.08675321443852511,
+    -0.02268254455026369
+   ]
+  },
+  {
+   "seed": 3,
+   "acc": 0.9412196914033799,
+   "equiv": 4917.07822373255,
+   "R": [
+    26944.758055139475,
+    24528.504861426587
+   ],
+   "delta_acc": [
+    0.0006429096252755327,
+    -0.005602498163115356,
+    0.007439382806759736
+   ],
+   "rel_cost": [
+    -0.04532867074642724,
+    -0.0759534618073302,
+    -0.013568378020935667
+   ]
+  }
+ ]
+}

analysis/bayes_sc_2026-10-25/v8/sc32_usd_cheapest.log

@@ -0,0 +1,45 @@
+2026-10-25 14:12:15,489 online_rl_train INFO: ThinkBooster math evaluator loaded
+reference agree2 (rotation pairs, -> q32b_think): acc 0.941 equiv 5151
+cost model usd: 1 equiv = 2e-07 USD; agree2 costs 0.00103 USD per problem
+  agree2                 acc 0.941 equiv   5151 (    +0; 0.00103 USD) | vs agree2 +0.0 [+0.0, +0.0]
+  SC@1                   acc 0.791 equiv   1150 ( -4001; 0.0002299 USD) | vs agree2 -15.0 [-16.3, -13.7]
+  SC@2                   acc 0.812 equiv   2299 ( -2851; 0.0004598 USD) | vs agree2 -12.9 [-14.2, -11.5]
+  SC@3                   acc 0.844 equiv   3449 ( -1702; 0.0006898 USD) | vs agree2 -9.6 [-10.9, -8.4]
+  SC@4                   acc 0.855 equiv   4598 (  -552; 0.0009197 USD) | vs agree2 -8.5 [-9.8, -7.3]
+  SC@5                   acc 0.863 equiv   5748 (  +597; 0.00115 USD) | vs agree2 -7.8 [-9.1, -6.5]
+  SC@6                   acc 0.869 equiv   6898 ( +1747; 0.00138 USD) | vs agree2 -7.1 [-8.4, -5.9]
+  SC@7                   acc 0.872 equiv   8047 ( +2897; 0.001609 USD) | vs agree2 -6.9 [-8.2, -5.6]
+  SC@8                   acc 0.876 equiv   9197 ( +4046; 0.001839 USD) | vs agree2 -6.4 [-7.8, -5.1]
+  AC(0.8)                acc 0.871 equiv   4412 (  -739; 0.0008824 USD) | vs agree2 -6.9 [-8.2, -5.6]
+  AC(0.8)+think@2        acc 0.941 equiv   5151 (    +0; 0.00103 USD) | vs agree2 +0.0 [+0.0, +0.0]
+  AC(0.8)+think@3        acc 0.940 equiv   5633 (  +482; 0.001127 USD) | vs agree2 -0.0 [-0.1, +0.0]
+  AC(0.8)+think@4        acc 0.934 equiv   5360 (  +210; 0.001072 USD) | vs agree2 -0.7 [-0.9, -0.4]
+  AC(0.9)                acc 0.876 equiv   6165 ( +1015; 0.001233 USD) | vs agree2 -6.5 [-7.8, -5.2]
+  AC(0.9)+think@2        acc 0.960 equiv   9755 ( +4605; 0.001951 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.9)+think@3        acc 0.954 equiv   6978 ( +1827; 0.001396 USD) | vs agree2 +1.4 [+1.1, +1.7]
+  AC(0.9)+think@4        acc 0.954 equiv   7566 ( +2415; 0.001513 USD) | vs agree2 +1.4 [+1.1, +1.6]
+  AC(0.95)               acc 0.876 equiv   7111 ( +1961; 0.001422 USD) | vs agree2 -6.5 [-7.8, -5.2]
+  AC(0.95)+think@2       acc 0.960 equiv   9755 ( +4605; 0.001951 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.95)+think@3       acc 0.960 equiv  10905 ( +5754; 0.002181 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.95)+think@4       acc 0.957 equiv   8507 ( +3356; 0.001701 USD) | vs agree2 +1.7 [+1.3, +2.1]
+  AC(0.99)               acc 0.876 equiv   8341 ( +3191; 0.001668 USD) | vs agree2 -6.4 [-7.8, -5.1]
+  AC(0.99)+think@2       acc 0.960 equiv   9755 ( +4605; 0.001951 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.99)+think@3       acc 0.960 equiv  10905 ( +5754; 0.002181 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.99)+think@4       acc 0.960 equiv  12054 ( +6904; 0.002411 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  always think           acc 0.960 equiv   7456 ( +2306; 0.001491 USD) | vs agree2 +2.0 [+1.3, +2.7]
+R    4000: acc 0.791 equiv   1151
+R    8000: acc 0.791 equiv   1151
+R   16000: acc 0.869 equiv   3078
+R   24000: acc 0.941 equiv   4937
+R   32000: acc 0.947 equiv   5097
+R   48000: acc 0.953 equiv   5360
+R   64000: acc 0.955 equiv   5518
+R   96000: acc 0.956 equiv   5646
+parity: adaptive-SC Bayes at agree2's cost acc 0.949 equiv 5180 (+29) | vs agree2 +0.9 [+0.3, +1.5] | R [44002, 30172] | {'think:q32b_think@0': 3576, 'stop@1': 3414, 'stop@2': 2949, 'think:q32b_think@2': 271, 'stop@3': 161, 'stop@5': 160, 'stop@4': 139, 'stop@7': 65}
+split 1: adaptive-SC Bayes at agree2's cost acc 0.944 equiv 5078 (-73) | vs agree2 +0.4 [-0.2, +1.0] | R [36742, 30333] | {'stop@1': 4215, 'think:q32b_think@0': 3688, 'stop@2': 2279, 'stop@4': 205, 'think:q32b_think@2': 104, 'think:q32b_think@3': 82, 'stop@7': 73, 'stop@5': 60}
+split 2: adaptive-SC Bayes at agree2's cost acc 0.947 equiv 5108 (-42) | vs agree2 +0.6 [+0.0, +1.2] | R [43715, 32840] | {'stop@1': 3926, 'think:q32b_think@0': 3600, 'stop@2': 2611, 'stop@5': 206, 'think:q32b_think@2': 152, 'stop@3': 119, 'think:q32b_think@3': 73, 'stop@7': 46}
+split 3: adaptive-SC Bayes at agree2's cost acc 0.948 equiv 5101 (-49) | vs agree2 +0.8 [+0.2, +1.4] | R [39018, 32701] | {'stop@1': 3963, 'think:q32b_think@0': 3648, 'stop@2': 2498, 'think:q32b_think@2': 258, 'stop@4': 169, 'stop@5': 99, 'stop@3': 77, 'think:q32b_think@3': 48}
+parity at agree2's accuracy: acc 0.941 cost 0.0009873 USD vs 0.00103 | cost -4.2% [-7.2, -0.9] | acc +0.0 [-0.6, +0.7]
+split 1 at agree2's accuracy: acc 0.941 cost 0.0009913 USD vs 0.00103 | cost -3.8% [-6.9, -0.6] | acc +0.1 [-0.6, +0.7]
+split 2 at agree2's accuracy: acc 0.939 cost 0.0009733 USD vs 0.00103 | cost -5.5% [-8.7, -2.3] | acc -0.2 [-0.8, +0.5]
+split 3 at agree2's accuracy: acc 0.941 cost 0.0009834 USD vs 0.00103 | cost -4.5% [-7.6, -1.4] | acc +0.1 [-0.6, +0.7]

analysis/bayes_sc_2026-10-25/v8/sc32_usd_cheapest.provenance.json

@@ -0,0 +1,92 @@
+{
+ "started_at": "2026-10-25T14:12:13+04:00",
+ "host": "ws006600",
+ "user": "vlad.smirnov",
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
+ "argv": [
+  "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
+  "--pool-dir",
+  "workdir/selector/pool_mth_k8_chat",
+  "--uhead",
+  "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+  "--task-prior",
+  "workdir/bayes/qprior/mth_qprior.json",
+  "--cost-model",
+  "usd",
+  "--price-scenario",
+  "cheapest",
+  "--tier",
+  "q32b_think=workdir/bayes/q32b_think_tier.json@qwen3-32b-fp8:think",
+  "--match-acc",
+  "--out",
+  "workdir/bayes/v8/sc32_usd_cheapest.json"
+ ],
+ "git": {
+  "commit": "f935c73993aae0c1137fdec2866aef2005007331",
+  "branch": "HEAD",
+  "subject": "Per-problem arrays in the Bayesian controllers; paired bootstrap against the hull mixture in currency_table",
+  "dirty": false,
+  "dirty_files": []
+ },
+ "gpus": {
+  "visible": "",
+  "devices": [
+   {
+    "index": "0",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
+   },
+   {
+    "index": "1",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
+   }
+  ]
+ },
+ "packages": {
+  "python": "3.11.15",
+  "torch": "2.9.1",
+  "vllm": "0.15.1",
+  "transformers": "4.57.6",
+  "speculators": "0.4.0.dev49",
+  "peft": null,
+  "accelerate": "1.13.0",
+  "numpy": "2.2.6",
+  "huggingface_hub": "0.36.2",
+  "openai": "2.32.0",
+  "wandb": "0.25.1"
+ },
+ "env": {
+  "CUDA_VISIBLE_DEVICES": ""
+ },
+ "models": [],
+ "files": {
+  "pool": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
+   "bytes": 65136887,
+   "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
+  },
+  "uhead": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+   "bytes": 9369632,
+   "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
+  },
+  "task_prior": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/qprior/mth_qprior.json",
+   "bytes": 28987,
+   "sha256": "897f405c643430dbe97023d9f8985e108330b82ed412a25ab2a4b84c992fbbd6"
+  },
+  "tier:q32b_think": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/q32b_think_tier.json",
+   "bytes": 26525,
+   "sha256": "4eef554b7fa4edff6cf11b63aadb44267d211bc25285eb260dcc1b45d2cf3a1f"
+  }
+ },
+ "finished_at": "2026-10-25T14:12:56+04:00",
+ "wall_secs": 43.6,
+ "n_problems": 1361
+}

analysis/bayes_sc_2026-10-25/v8/sc32_usd_dear_32b.arrays.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5095d5f10cc996fa92e5a2ab199e4a7618adb98f77e7d2b452c5a1e4b5fed127
+size 194848

analysis/bayes_sc_2026-10-25/v8/sc32_usd_dear_32b.json

@@ -0,0 +1,494 @@
+{
+ "ref": {
+  "acc": 0.9405767817781043,
+  "equiv": 8100.071643093312
+ },
+ "cost_model": {
+  "currency": "usd",
+  "unit": "USD",
+  "per_equiv": 2.0000000000000002e-07,
+  "scenario": "dear_32b",
+  "base": "qwen3-8b:nothink"
+ },
+ "baselines": {
+  "agree2": {
+   "acc": 0.9405767817781043,
+   "equiv": 8100.071643093312,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "SC@1": {
+   "acc": 0.7907788390889052,
+   "equiv": 1149.6065393093313,
+   "delta_vs_agree2": [
+    -0.14979794268919913,
+    -0.16256429096252756,
+    -0.1369374540778839
+   ]
+  },
+  "SC@2": {
+   "acc": 0.8118111682586333,
+   "equiv": 2299.2130786186626,
+   "delta_vs_agree2": [
+    -0.12876561351947097,
+    -0.14226671565025717,
+    -0.11544819985304923
+   ]
+  },
+  "SC@3": {
+   "acc": 0.8444158706833211,
+   "equiv": 3448.819617927994,
+   "delta_vs_agree2": [
+    -0.09616091109478325,
+    -0.10911324393828067,
+    -0.08357825128581925
+   ]
+  },
+  "SC@4": {
+   "acc": 0.8553453343130052,
+   "equiv": 4598.426157237325,
+   "delta_vs_agree2": [
+    -0.0852314474650992,
+    -0.09809652828802351,
+    -0.07274063188831742
+   ]
+  },
+  "SC@5": {
+   "acc": 0.8630602498163116,
+   "equiv": 5748.032696546657,
+   "delta_vs_agree2": [
+    -0.0775165319617928,
+    -0.09056070903747244,
+    -0.06493387215282881
+   ]
+  },
+  "SC@6": {
+   "acc": 0.8693056576047025,
+   "equiv": 6897.639235855988,
+   "delta_vs_agree2": [
+    -0.07127112417340191,
+    -0.08440714548126378,
+    -0.05905584129316679
+   ]
+  },
+  "SC@7": {
+   "acc": 0.8717854518736223,
+   "equiv": 8047.245775165319,
+   "delta_vs_agree2": [
+    -0.068791329904482,
+    -0.08174136664217486,
+    -0.05648190668626012
+   ]
+  },
+  "SC@8": {
+   "acc": 0.8761021307861866,
+   "equiv": 9196.85231447465,
+   "delta_vs_agree2": [
+    -0.06447465099191771,
+    -0.07788850110213079,
+    -0.05143277002204261
+   ]
+  },
+  "AC(0.8)": {
+   "acc": 0.8714180749448934,
+   "equiv": 4411.8378260470245,
+   "delta_vs_agree2": [
+    -0.06915870683321088,
+    -0.08238886847905952,
+    -0.05639235855988244
+   ]
+  },
+  "AC(0.8)+think@2": {
+   "acc": 0.9405767817781043,
+   "equiv": 8100.071643093312,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@3": {
+   "acc": 0.9403930933137399,
+   "equiv": 8579.525822005875,
+   "delta_vs_agree2": [
+    -0.0001836884643644379,
+    -0.0005510653930933138,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@4": {
+   "acc": 0.9340558412931668,
+   "equiv": 7525.915664033799,
+   "delta_vs_agree2": [
+    -0.006520940484937546,
+    -0.009184423218221896,
+    -0.004132990448199853
+   ]
+  },
+  "AC(0.9)": {
+   "acc": 0.8757347538574578,
+   "equiv": 6165.147180382072,
+   "delta_vs_agree2": [
+    -0.06484202792064658,
+    -0.07834542615723733,
+    -0.05198383541513593
+   ]
+  },
+  "AC(0.9)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 17466.341513592943,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.9)+think@3": {
+   "acc": 0.9542615723732549,
+   "equiv": 10628.620187362234,
+   "delta_vs_agree2": [
+    0.013684790595150624,
+    0.010929463629684055,
+    0.016623806024981633
+   ]
+  },
+  "AC(0.9)+think@4": {
+   "acc": 0.9540778839088905,
+   "equiv": 11214.049223916236,
+   "delta_vs_agree2": [
+    0.013501102130786187,
+    0.010745775165319618,
+    0.0162587252020573
+   ]
+  },
+  "AC(0.95)": {
+   "acc": 0.8760102865540044,
+   "equiv": 7111.373874908156,
+   "delta_vs_agree2": [
+    -0.06456649522409992,
+    -0.07806989346069067,
+    -0.05152461425422483
+   ]
+  },
+  "AC(0.95)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 17466.341513592943,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.95)+think@3": {
+   "acc": 0.9603232916972814,
+   "equiv": 18615.948052902277,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.95)+think@4": {
+   "acc": 0.9572005878030859,
+   "equiv": 12549.804578434974,
+   "delta_vs_agree2": [
+    0.016623806024981633,
+    0.012950036737692874,
+    0.020575404114621592
+   ]
+  },
+  "AC(0.99)": {
+   "acc": 0.8761021307861866,
+   "equiv": 8341.35433504776,
+   "delta_vs_agree2": [
+    -0.06447465099191771,
+    -0.07788850110213079,
+    -0.05143277002204261
+   ]
+  },
+  "AC(0.99)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 17466.341513592943,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.99)+think@3": {
+   "acc": 0.9603232916972814,
+   "equiv": 18615.948052902277,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.99)+think@4": {
+   "acc": 0.9603232916972814,
+   "equiv": 19765.554592211607,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "always think": {
+   "acc": 0.9603232916972814,
+   "equiv": 15167.128434974282,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  }
+ },
+ "frontier": [
+  {
+   "R": 4000,
+   "acc": 0.7907788390889052,
+   "equiv": 1150.5956557678178
+  },
+  {
+   "R": 8000,
+   "acc": 0.7907788390889052,
+   "equiv": 1150.5956557678178
+  },
+  {
+   "R": 16000,
+   "acc": 0.8075863335782513,
+   "equiv": 1396.3716017634092
+  },
+  {
+   "R": 24000,
+   "acc": 0.8149338721528288,
+   "equiv": 1676.2500183688464
+  },
+  {
+   "R": 32000,
+   "acc": 0.901451138868479,
+   "equiv": 5596.907228141072
+  },
+  {
+   "R": 48000,
+   "acc": 0.9304739162380602,
+   "equiv": 7248.4472033431275
+  },
+  {
+   "R": 64000,
+   "acc": 0.9386480529022777,
+   "equiv": 7845.916568699485
+  },
+  {
+   "R": 96000,
+   "acc": 0.9466385011021308,
+   "equiv": 9170.761462160175
+  }
+ ],
+ "matched": [
+  {
+   "seed": 0,
+   "acc": 0.9389235855988244,
+   "equiv": 7960.538997060984,
+   "R": [
+    65162.81182746214,
+    87240.34686240471
+   ],
+   "delta_vs_agree2": [
+    -0.0016531961792799412,
+    -0.005602498163115356,
+    0.002112417340191036
+   ],
+   "actions": {
+    "stop@2": 7205,
+    "stop@4": 420,
+    "stop@5": 435,
+    "stop@6": 130,
+    "stop@7": 134,
+    "think:q32b_think@6": 95,
+    "stop@3": 613,
+    "think:q32b_think@2": 633,
+    "think:q32b_think@3": 288,
+    "think:q32b_think@4": 489,
+    "stop@8": 172,
+    "think:q32b_think@7": 30,
+    "think:q32b_think@5": 196,
+    "think:q32b_think@1": 1,
+    "think:q32b_think@8": 47
+   }
+  },
+  {
+   "seed": 1,
+   "acc": 0.9402094048493754,
+   "equiv": 7959.679114621601,
+   "R": [
+    66863.9667712638,
+    89949.39614832334
+   ],
+   "delta_vs_agree2": [
+    -0.0003673769287288758,
+    -0.004041146216017634,
+    0.0034900808229243203
+   ],
+   "actions": {
+    "stop@8": 132,
+    "stop@7": 141,
+    "stop@6": 154,
+    "stop@4": 273,
+    "stop@3": 773,
+    "stop@2": 7096,
+    "think:q32b_think@3": 459,
+    "stop@5": 491,
+    "think:q32b_think@4": 225,
+    "think:q32b_think@2": 817,
+    "think:q32b_think@7": 67,
+    "think:q32b_think@6": 133,
+    "think:q32b_think@5": 89,
+    "think:q32b_think@8": 37,
+    "think:q32b_think@1": 1
+   }
+  },
+  {
+   "seed": 2,
+   "acc": 0.9414952240999265,
+   "equiv": 8091.537027002203,
+   "R": [
+    70038.00058507305,
+    76742.53748397072
+   ],
+   "delta_vs_agree2": [
+    0.0009184423218221896,
+    -0.0030308596620132254,
+    0.004775900073475386
+   ],
+   "actions": {
+    "stop@5": 587,
+    "stop@8": 164,
+    "stop@7": 117,
+    "stop@6": 145,
+    "stop@4": 390,
+    "stop@3": 508,
+    "stop@2": 7112,
+    "think:q32b_think@6": 71,
+    "think:q32b_think@2": 611,
+    "think:q32b_think@3": 686,
+    "think:q32b_think@4": 177,
+    "think:q32b_think@5": 205,
+    "think:q32b_think@8": 76,
+    "think:q32b_think@7": 39
+   }
+  },
+  {
+   "seed": 3,
+   "acc": 0.9407604702424688,
+   "equiv": 8151.956695444526,
+   "R": [
+    95584.05852332129,
+    55620.8972776458
+   ],
+   "delta_vs_agree2": [
+    0.0001836884643644379,
+    -0.004041146216017634,
+    0.004224834680382072
+   ],
+   "actions": {
+    "stop@2": 6370,
+    "stop@6": 209,
+    "think:q32b_think@3": 639,
+    "stop@5": 710,
+    "stop@3": 1259,
+    "stop@4": 231,
+    "think:q32b_think@6": 37,
+    "think:q32b_think@5": 183,
+    "stop@8": 98,
+    "stop@7": 122,
+    "think:q32b_think@7": 52,
+    "think:q32b_think@8": 35,
+    "think:q32b_think@4": 213,
+    "think:q32b_think@2": 728,
+    "think:q32b_think@1": 2
+   }
+  }
+ ],
+ "matched_acc": [
+  {
+   "seed": 0,
+   "acc": 0.9443423952975754,
+   "equiv": 8886.334340558411,
+   "R": [
+    65163.05893015895,
+    59330.632516829646
+   ],
+   "delta_acc": [
+    0.003765613519470977,
+    -0.0011044268919911829,
+    0.008633357825128581
+   ],
+   "rel_cost": [
+    0.09706861026785137,
+    0.06353250901962904,
+    0.13124616628486666
+   ]
+  },
+  {
+   "seed": 1,
+   "acc": 0.9376377663482733,
+   "equiv": 7669.243277002204,
+   "R": [
+    53842.244973533416,
+    58447.950800650295
+   ],
+   "delta_acc": [
+    -0.0029390154298310064,
+    -0.00707200587803086,
+    0.0011939750183688465
+   ],
+   "rel_cost": [
+    -0.05318821673120184,
+    -0.07677697942878778,
+    -0.030730151797483382
+   ]
+  },
+  {
+   "seed": 2,
+   "acc": 0.9376377663482733,
+   "equiv": 7766.884827332842,
+   "R": [
+    57581.5891689974,
+    61119.50696680797
+   ],
+   "delta_acc": [
+    -0.0029390154298310064,
+    -0.007255694342395298,
+    0.0013776634827332844
+   ],
+   "rel_cost": [
+    -0.041133810963828746,
+    -0.06676285007550273,
+    -0.014911697390608412
+   ]
+  },
+  {
+   "seed": 3,
+   "acc": 0.9388317413666422,
+   "equiv": 7926.542427443055,
+   "R": [
+    67255.20709236493,
+    52691.11102712724
+   ],
+   "delta_acc": [
+    -0.0017450404114621601,
+    -0.006061719324026451,
+    0.002296105804555474
+   ],
+   "rel_cost": [
+    -0.021423170472599472,
+    -0.045324401010637935,
+    0.0024378930119188617
+   ]
+  }
+ ]
+}

analysis/bayes_sc_2026-10-25/v8/sc32_usd_dear_32b.log

@@ -0,0 +1,45 @@
+2026-10-25 14:12:15,505 online_rl_train INFO: ThinkBooster math evaluator loaded
+reference agree2 (rotation pairs, -> q32b_think): acc 0.941 equiv 8100
+cost model usd: 1 equiv = 2e-07 USD; agree2 costs 0.00162 USD per problem
+  agree2                 acc 0.941 equiv   8100 (    +0; 0.00162 USD) | vs agree2 +0.0 [+0.0, +0.0]
+  SC@1                   acc 0.791 equiv   1150 ( -6950; 0.0002299 USD) | vs agree2 -15.0 [-16.3, -13.7]
+  SC@2                   acc 0.812 equiv   2299 ( -5801; 0.0004598 USD) | vs agree2 -12.9 [-14.2, -11.5]
+  SC@3                   acc 0.844 equiv   3449 ( -4651; 0.0006898 USD) | vs agree2 -9.6 [-10.9, -8.4]
+  SC@4                   acc 0.855 equiv   4598 ( -3502; 0.0009197 USD) | vs agree2 -8.5 [-9.8, -7.3]
+  SC@5                   acc 0.863 equiv   5748 ( -2352; 0.00115 USD) | vs agree2 -7.8 [-9.1, -6.5]
+  SC@6                   acc 0.869 equiv   6898 ( -1202; 0.00138 USD) | vs agree2 -7.1 [-8.4, -5.9]
+  SC@7                   acc 0.872 equiv   8047 (   -53; 0.001609 USD) | vs agree2 -6.9 [-8.2, -5.6]
+  SC@8                   acc 0.876 equiv   9197 ( +1097; 0.001839 USD) | vs agree2 -6.4 [-7.8, -5.1]
+  AC(0.8)                acc 0.871 equiv   4412 ( -3688; 0.0008824 USD) | vs agree2 -6.9 [-8.2, -5.6]
+  AC(0.8)+think@2        acc 0.941 equiv   8100 (    +0; 0.00162 USD) | vs agree2 +0.0 [+0.0, +0.0]
+  AC(0.8)+think@3        acc 0.940 equiv   8580 (  +479; 0.001716 USD) | vs agree2 -0.0 [-0.1, +0.0]
+  AC(0.8)+think@4        acc 0.934 equiv   7526 (  -574; 0.001505 USD) | vs agree2 -0.7 [-0.9, -0.4]
+  AC(0.9)                acc 0.876 equiv   6165 ( -1935; 0.001233 USD) | vs agree2 -6.5 [-7.8, -5.2]
+  AC(0.9)+think@2        acc 0.960 equiv  17466 ( +9366; 0.003493 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.9)+think@3        acc 0.954 equiv  10629 ( +2529; 0.002126 USD) | vs agree2 +1.4 [+1.1, +1.7]
+  AC(0.9)+think@4        acc 0.954 equiv  11214 ( +3114; 0.002243 USD) | vs agree2 +1.4 [+1.1, +1.6]
+  AC(0.95)               acc 0.876 equiv   7111 (  -989; 0.001422 USD) | vs agree2 -6.5 [-7.8, -5.2]
+  AC(0.95)+think@2       acc 0.960 equiv  17466 ( +9366; 0.003493 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.95)+think@3       acc 0.960 equiv  18616 (+10516; 0.003723 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.95)+think@4       acc 0.957 equiv  12550 ( +4450; 0.00251 USD) | vs agree2 +1.7 [+1.3, +2.1]
+  AC(0.99)               acc 0.876 equiv   8341 (  +241; 0.001668 USD) | vs agree2 -6.4 [-7.8, -5.1]
+  AC(0.99)+think@2       acc 0.960 equiv  17466 ( +9366; 0.003493 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.99)+think@3       acc 0.960 equiv  18616 (+10516; 0.003723 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.99)+think@4       acc 0.960 equiv  19766 (+11665; 0.003953 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  always think           acc 0.960 equiv  15167 ( +7067; 0.003033 USD) | vs agree2 +2.0 [+1.3, +2.7]
+R    4000: acc 0.791 equiv   1151
+R    8000: acc 0.791 equiv   1151
+R   16000: acc 0.808 equiv   1396
+R   24000: acc 0.815 equiv   1676
+R   32000: acc 0.901 equiv   5597
+R   48000: acc 0.930 equiv   7248
+R   64000: acc 0.939 equiv   7846
+R   96000: acc 0.947 equiv   9171
+parity: adaptive-SC Bayes at agree2's cost acc 0.939 equiv 7961 (-140) | vs agree2 -0.2 [-0.6, +0.2] | R [65163, 87240] | {'stop@2': 7205, 'think:q32b_think@2': 633, 'stop@3': 613, 'think:q32b_think@4': 489, 'stop@5': 435, 'stop@4': 420, 'think:q32b_think@3': 288, 'think:q32b_think@5': 196}
+split 1: adaptive-SC Bayes at agree2's cost acc 0.940 equiv 7960 (-140) | vs agree2 -0.0 [-0.4, +0.3] | R [66864, 89949] | {'stop@2': 7096, 'think:q32b_think@2': 817, 'stop@3': 773, 'stop@5': 491, 'think:q32b_think@3': 459, 'stop@4': 273, 'think:q32b_think@4': 225, 'stop@6': 154}
+split 2: adaptive-SC Bayes at agree2's cost acc 0.941 equiv 8092 (-9) | vs agree2 +0.1 [-0.3, +0.5] | R [70038, 76743] | {'stop@2': 7112, 'think:q32b_think@3': 686, 'think:q32b_think@2': 611, 'stop@5': 587, 'stop@3': 508, 'stop@4': 390, 'think:q32b_think@5': 205, 'think:q32b_think@4': 177}
+split 3: adaptive-SC Bayes at agree2's cost acc 0.941 equiv 8152 (+52) | vs agree2 +0.0 [-0.4, +0.4] | R [95584, 55621] | {'stop@2': 6370, 'stop@3': 1259, 'think:q32b_think@2': 728, 'stop@5': 710, 'think:q32b_think@3': 639, 'stop@4': 231, 'think:q32b_think@4': 213, 'stop@6': 209}
+parity at agree2's accuracy: acc 0.944 cost 0.001777 USD vs 0.00162 | cost +9.7% [+6.4, +13.1] | acc +0.4 [-0.1, +0.9]
+split 1 at agree2's accuracy: acc 0.938 cost 0.001534 USD vs 0.00162 | cost -5.3% [-7.7, -3.1] | acc -0.3 [-0.7, +0.1]
+split 2 at agree2's accuracy: acc 0.938 cost 0.001553 USD vs 0.00162 | cost -4.1% [-6.7, -1.5] | acc -0.3 [-0.7, +0.1]
+split 3 at agree2's accuracy: acc 0.939 cost 0.001585 USD vs 0.00162 | cost -2.1% [-4.5, +0.2] | acc -0.2 [-0.6, +0.2]

analysis/bayes_sc_2026-10-25/v8/sc32_usd_dear_32b.provenance.json

@@ -0,0 +1,92 @@
+{
+ "started_at": "2026-10-25T14:12:13+04:00",
+ "host": "ws006600",
+ "user": "vlad.smirnov",
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
+ "argv": [
+  "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
+  "--pool-dir",
+  "workdir/selector/pool_mth_k8_chat",
+  "--uhead",
+  "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+  "--task-prior",
+  "workdir/bayes/qprior/mth_qprior.json",
+  "--cost-model",
+  "usd",
+  "--price-scenario",
+  "dear_32b",
+  "--tier",
+  "q32b_think=workdir/bayes/q32b_think_tier.json@qwen3-32b-fp8:think",
+  "--match-acc",
+  "--out",
+  "workdir/bayes/v8/sc32_usd_dear_32b.json"
+ ],
+ "git": {
+  "commit": "f935c73993aae0c1137fdec2866aef2005007331",
+  "branch": "HEAD",
+  "subject": "Per-problem arrays in the Bayesian controllers; paired bootstrap against the hull mixture in currency_table",
+  "dirty": false,
+  "dirty_files": []
+ },
+ "gpus": {
+  "visible": "",
+  "devices": [
+   {
+    "index": "0",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
+   },
+   {
+    "index": "1",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
+   }
+  ]
+ },
+ "packages": {
+  "python": "3.11.15",
+  "torch": "2.9.1",
+  "vllm": "0.15.1",
+  "transformers": "4.57.6",
+  "speculators": "0.4.0.dev49",
+  "peft": null,
+  "accelerate": "1.13.0",
+  "numpy": "2.2.6",
+  "huggingface_hub": "0.36.2",
+  "openai": "2.32.0",
+  "wandb": "0.25.1"
+ },
+ "env": {
+  "CUDA_VISIBLE_DEVICES": ""
+ },
+ "models": [],
+ "files": {
+  "pool": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
+   "bytes": 65136887,
+   "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
+  },
+  "uhead": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+   "bytes": 9369632,
+   "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
+  },
+  "task_prior": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/qprior/mth_qprior.json",
+   "bytes": 28987,
+   "sha256": "897f405c643430dbe97023d9f8985e108330b82ed412a25ab2a4b84c992fbbd6"
+  },
+  "tier:q32b_think": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/q32b_think_tier.json",
+   "bytes": 26525,
+   "sha256": "4eef554b7fa4edff6cf11b63aadb44267d211bc25285eb260dcc1b45d2cf3a1f"
+  }
+ },
+ "finished_at": "2026-10-25T14:13:08+04:00",
+ "wall_secs": 55.6,
+ "n_problems": 1361
+}

analysis/bayes_sc_2026-10-25/v8/sc32_usd_openrouter_only.arrays.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1dbe6f2b9b039025e4a849c82bbfb3d8c9096cd4cce882bfac7175f654b5258c
+size 251070

analysis/bayes_sc_2026-10-25/v8/sc32_usd_openrouter_only.json

@@ -0,0 +1,499 @@
+{
+ "ref": {
+  "acc": 0.9405767817781043,
+  "equiv": 3553.955527206078
+ },
+ "cost_model": {
+  "currency": "usd",
+  "unit": "USD",
+  "per_equiv": 4.5500000000000004e-07,
+  "scenario": "openrouter_only",
+  "base": "qwen3-8b:nothink"
+ },
+ "baselines": {
+  "agree2": {
+   "acc": 0.9405767817781043,
+   "equiv": 3553.955527206078,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "SC@1": {
+   "acc": 0.7907788390889052,
+   "equiv": 1150.3112732234702,
+   "delta_vs_agree2": [
+    -0.14979794268919913,
+    -0.16256429096252756,
+    -0.1369374540778839
+   ]
+  },
+  "SC@2": {
+   "acc": 0.8118111682586333,
+   "equiv": 2300.6225464469403,
+   "delta_vs_agree2": [
+    -0.12876561351947097,
+    -0.14226671565025717,
+    -0.11544819985304923
+   ]
+  },
+  "SC@3": {
+   "acc": 0.8444158706833211,
+   "equiv": 3450.9338196704107,
+   "delta_vs_agree2": [
+    -0.09616091109478325,
+    -0.10911324393828067,
+    -0.08357825128581925
+   ]
+  },
+  "SC@4": {
+   "acc": 0.8553453343130052,
+   "equiv": 4601.245092893881,
+   "delta_vs_agree2": [
+    -0.0852314474650992,
+    -0.09809652828802351,
+    -0.07274063188831742
+   ]
+  },
+  "SC@5": {
+   "acc": 0.8630602498163116,
+   "equiv": 5751.556366117351,
+   "delta_vs_agree2": [
+    -0.0775165319617928,
+    -0.09056070903747244,
+    -0.06493387215282881
+   ]
+  },
+  "SC@6": {
+   "acc": 0.8693056576047025,
+   "equiv": 6901.867639340821,
+   "delta_vs_agree2": [
+    -0.07127112417340191,
+    -0.08440714548126378,
+    -0.05905584129316679
+   ]
+  },
+  "SC@7": {
+   "acc": 0.8717854518736223,
+   "equiv": 8052.178912564291,
+   "delta_vs_agree2": [
+    -0.068791329904482,
+    -0.08174136664217486,
+    -0.05648190668626012
+   ]
+  },
+  "SC@8": {
+   "acc": 0.8761021307861866,
+   "equiv": 9202.490185787761,
+   "delta_vs_agree2": [
+    -0.06447465099191771,
+    -0.07788850110213079,
+    -0.05143277002204261
+   ]
+  },
+  "AC(0.8)": {
+   "acc": 0.8714180749448934,
+   "equiv": 4414.206011861026,
+   "delta_vs_agree2": [
+    -0.06915870683321088,
+    -0.08238886847905952,
+    -0.05639235855988244
+   ]
+  },
+  "AC(0.8)+think@2": {
+   "acc": 0.9405767817781043,
+   "equiv": 3553.955527206078,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@3": {
+   "acc": 0.9403930933137399,
+   "equiv": 4037.685613761697,
+   "delta_vs_agree2": [
+    -0.0001836884643644379,
+    -0.0005510653930933138,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@4": {
+   "acc": 0.9340558412931668,
+   "equiv": 4189.212759485188,
+   "delta_vs_agree2": [
+    -0.006520940484937546,
+    -0.009184423218221896,
+    -0.004132990448199853
+   ]
+  },
+  "AC(0.9)": {
+   "acc": 0.8757347538574578,
+   "equiv": 6168.549855673349,
+   "delta_vs_agree2": [
+    -0.06484202792064658,
+    -0.07834542615723733,
+    -0.05198383541513593
+   ]
+  },
+  "AC(0.9)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 5578.010762932879,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.9)+think@3": {
+   "acc": 0.9542615723732549,
+   "equiv": 5002.208335015463,
+   "delta_vs_agree2": [
+    0.013684790595150624,
+    0.010929463629684055,
+    0.016623806024981633
+   ]
+  },
+  "AC(0.9)+think@4": {
+   "acc": 0.9540778839088905,
+   "equiv": 5591.969512155735,
+   "delta_vs_agree2": [
+    0.013501102130786187,
+    0.010745775165319618,
+    0.0162587252020573
+   ]
+  },
+  "AC(0.95)": {
+   "acc": 0.8760102865540044,
+   "equiv": 7115.399391203946,
+   "delta_vs_agree2": [
+    -0.06456649522409992,
+    -0.07806989346069067,
+    -0.05152461425422483
+   ]
+  },
+  "AC(0.95)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 5578.010762932879,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.95)+think@3": {
+   "acc": 0.9603232916972814,
+   "equiv": 6728.322036156349,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.95)+think@4": {
+   "acc": 0.9572005878030859,
+   "equiv": 6319.233797062599,
+   "delta_vs_agree2": [
+    0.016623806024981633,
+    0.012950036737692874,
+    0.020575404114621592
+   ]
+  },
+  "AC(0.99)": {
+   "acc": 0.8761021307861866,
+   "equiv": 8346.294659913927,
+   "delta_vs_agree2": [
+    -0.06447465099191771,
+    -0.07788850110213079,
+    -0.05143277002204261
+   ]
+  },
+  "AC(0.99)+think@2": {
+   "acc": 0.9603232916972814,
+   "equiv": 5578.010762932879,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.99)+think@3": {
+   "acc": 0.9603232916972814,
+   "equiv": 6728.322036156349,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "AC(0.99)+think@4": {
+   "acc": 0.9603232916972814,
+   "equiv": 7878.633309379819,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  },
+  "always think": {
+   "acc": 0.9603232916972814,
+   "equiv": 3277.388216485939,
+   "delta_vs_agree2": [
+    0.019746509919177077,
+    0.012766348273328435,
+    0.026726671565025718
+   ]
+  }
+ },
+ "frontier": [
+  {
+   "R": 4000,
+   "acc": 0.8467119764878765,
+   "equiv": 1677.7417610677344
+  },
+  {
+   "R": 8000,
+   "acc": 0.9099008082292432,
+   "equiv": 2190.841744717443
+  },
+  {
+   "R": 16000,
+   "acc": 0.9398420279206466,
+   "equiv": 2525.680790021881
+  },
+  {
+   "R": 24000,
+   "acc": 0.945077149155033,
+   "equiv": 2701.0134649700044
+  },
+  {
+   "R": 32000,
+   "acc": 0.9544452608376194,
+   "equiv": 2880.1967099579333
+  },
+  {
+   "R": 48000,
+   "acc": 0.9566495224099927,
+   "equiv": 3002.5153274902905
+  },
+  {
+   "R": 64000,
+   "acc": 0.9584864070536371,
+   "equiv": 3126.361602247862
+  },
+  {
+   "R": 96000,
+   "acc": 0.959313005143277,
+   "equiv": 3268.4759458946637
+  }
+ ],
+ "matched": [
+  {
+   "seed": 0,
+   "acc": 0.9579353416605437,
+   "equiv": 3528.2999255153372,
+   "R": [
+    182332.03982533354,
+    232322.57610077015
+   ],
+   "delta_vs_agree2": [
+    0.01735855988243938,
+    0.012123438648052902,
+    0.02305290227773696
+   ],
+   "actions": {
+    "stop@2": 3251,
+    "stop@3": 1856,
+    "stop@5": 282,
+    "think:q32b_think@3": 121,
+    "think:q32b_think@0": 3576,
+    "stop@4": 664,
+    "stop@6": 62,
+    "think:q32b_think@8": 31,
+    "think:q32b_think@7": 45,
+    "think:q32b_think@6": 38,
+    "think:q32b_think@5": 116,
+    "stop@8": 31,
+    "stop@7": 56,
+    "think:q32b_think@4": 35,
+    "think:q32b_think@1": 649,
+    "think:q32b_think@2": 75
+   }
+  },
+  {
+   "seed": 1,
+   "acc": 0.9581190301249082,
+   "equiv": 4443.948748698032,
+   "R": [
+    377647.0835837676,
+    256298.70147382346
+   ],
+   "delta_vs_agree2": [
+    0.01754224834680382,
+    0.011204996326230713,
+    0.024063188831741366
+   ],
+   "actions": {
+    "think:q32b_think@0": 3984,
+    "think:q32b_think@8": 358,
+    "think:q32b_think@1": 1039,
+    "stop@2": 3368,
+    "stop@5": 169,
+    "stop@3": 426,
+    "think:q32b_think@3": 141,
+    "think:q32b_think@2": 230,
+    "think:q32b_think@4": 161,
+    "think:q32b_think@5": 119,
+    "think:q32b_think@6": 59,
+    "think:q32b_think@7": 46,
+    "stop@4": 650,
+    "stop@7": 104,
+    "stop@8": 10,
+    "stop@6": 24
+   }
+  },
+  {
+   "seed": 2,
+   "acc": 0.9585782512858193,
+   "equiv": 3644.2933288790564,
+   "R": [
+    195464.0125687913,
+    172167.85305362387
+   ],
+   "delta_vs_agree2": [
+    0.018001469507714914,
+    0.01230712711241734,
+    0.024155033063923585
+   ],
+   "actions": {
+    "think:q32b_think@2": 268,
+    "think:q32b_think@3": 254,
+    "stop@7": 33,
+    "stop@6": 46,
+    "stop@5": 237,
+    "stop@4": 629,
+    "stop@3": 2999,
+    "stop@2": 1498,
+    "think:q32b_think@4": 56,
+    "think:q32b_think@0": 3600,
+    "think:q32b_think@7": 11,
+    "stop@8": 58,
+    "think:q32b_think@5": 34,
+    "think:q32b_think@6": 30,
+    "think:q32b_think@8": 17,
+    "think:q32b_think@1": 1118
+   }
+  },
+  {
+   "seed": 3,
+   "acc": 0.9570168993387216,
+   "equiv": 3346.469277801552,
+   "R": [
+    176829.00853116572,
+    499998.353065514
+   ],
+   "delta_vs_agree2": [
+    0.016440117560617192,
+    0.010102865540044085,
+    0.02286921381337252
+   ],
+   "actions": {
+    "stop@2": 2492,
+    "stop@3": 1572,
+    "think:q32b_think@1": 870,
+    "stop@5": 206,
+    "think:q32b_think@0": 5448,
+    "think:q32b_think@3": 13,
+    "stop@6": 39,
+    "stop@8": 23,
+    "stop@7": 49,
+    "think:q32b_think@8": 32,
+    "think:q32b_think@2": 66,
+    "think:q32b_think@5": 8,
+    "think:q32b_think@4": 14,
+    "think:q32b_think@7": 13,
+    "think:q32b_think@6": 3,
+    "stop@4": 40
+   }
+  }
+ ],
+ "matched_acc": [
+  {
+   "seed": 0,
+   "acc": 0.9412196914033799,
+   "equiv": 2580.725200846178,
+   "R": [
+    21906.703418265966,
+    11092.082736867853
+   ],
+   "delta_acc": [
+    0.0006429096252755327,
+    -0.005696638501102131,
+    0.007349834680382059
+   ],
+   "rel_cost": [
+    -0.2738442613897872,
+    -0.3008419126634079,
+    -0.24720560779686943
+   ]
+  },
+  {
+   "seed": 1,
+   "acc": 0.9420462894930198,
+   "equiv": 2564.7022765258253,
+   "R": [
+    20289.73040181801,
+    18849.498095276507
+   ],
+   "delta_acc": [
+    0.0014695077149155032,
+    -0.004778196179279941,
+    0.008268277002204248
+   ],
+   "rel_cost": [
+    -0.2783527377051757,
+    -0.305658830947544,
+    -0.2520288138861257
+   ]
+  },
+  {
+   "seed": 2,
+   "acc": 0.9416789125642909,
+   "equiv": 2563.2277248467917,
+   "R": [
+    20103.10371473359,
+    18384.193553033187
+   ],
+   "delta_acc": [
+    0.0011021307861866275,
+    -0.005143277002204262,
+    0.007533523144746497
+   ],
+   "rel_cost": [
+    -0.27876764207517857,
+    -0.30600929613142014,
+    -0.25315713056667916
+   ]
+  },
+  {
+   "seed": 3,
+   "acc": 0.9428728875826599,
+   "equiv": 2611.7601062970825,
+   "R": [
+    11205.398916819575,
+    19744.7040072149
+   ],
+   "delta_acc": [
+    0.002296105804555474,
+    -0.0036760653930933136,
+    0.008449669360764144
+   ],
+   "rel_cost": [
+    -0.2651117645385105,
+    -0.2921360212478392,
+    -0.24018601935296416
+   ]
+  }
+ ]
+}

analysis/bayes_sc_2026-10-25/v8/sc32_usd_openrouter_only.log

@@ -0,0 +1,45 @@
+2026-10-25 14:12:15,532 online_rl_train INFO: ThinkBooster math evaluator loaded
+reference agree2 (rotation pairs, -> q32b_think): acc 0.941 equiv 3554
+cost model usd: 1 equiv = 4.55e-07 USD; agree2 costs 0.001617 USD per problem
+  agree2                 acc 0.941 equiv   3554 (    +0; 0.001617 USD) | vs agree2 +0.0 [+0.0, +0.0]
+  SC@1                   acc 0.791 equiv   1150 ( -2404; 0.0005234 USD) | vs agree2 -15.0 [-16.3, -13.7]
+  SC@2                   acc 0.812 equiv   2301 ( -1253; 0.001047 USD) | vs agree2 -12.9 [-14.2, -11.5]
+  SC@3                   acc 0.844 equiv   3451 (  -103; 0.00157 USD) | vs agree2 -9.6 [-10.9, -8.4]
+  SC@4                   acc 0.855 equiv   4601 ( +1047; 0.002094 USD) | vs agree2 -8.5 [-9.8, -7.3]
+  SC@5                   acc 0.863 equiv   5752 ( +2198; 0.002617 USD) | vs agree2 -7.8 [-9.1, -6.5]
+  SC@6                   acc 0.869 equiv   6902 ( +3348; 0.00314 USD) | vs agree2 -7.1 [-8.4, -5.9]
+  SC@7                   acc 0.872 equiv   8052 ( +4498; 0.003664 USD) | vs agree2 -6.9 [-8.2, -5.6]
+  SC@8                   acc 0.876 equiv   9202 ( +5649; 0.004187 USD) | vs agree2 -6.4 [-7.8, -5.1]
+  AC(0.8)                acc 0.871 equiv   4414 (  +860; 0.002008 USD) | vs agree2 -6.9 [-8.2, -5.6]
+  AC(0.8)+think@2        acc 0.941 equiv   3554 (    +0; 0.001617 USD) | vs agree2 +0.0 [+0.0, +0.0]
+  AC(0.8)+think@3        acc 0.940 equiv   4038 (  +484; 0.001837 USD) | vs agree2 -0.0 [-0.1, +0.0]
+  AC(0.8)+think@4        acc 0.934 equiv   4189 (  +635; 0.001906 USD) | vs agree2 -0.7 [-0.9, -0.4]
+  AC(0.9)                acc 0.876 equiv   6169 ( +2615; 0.002807 USD) | vs agree2 -6.5 [-7.8, -5.2]
+  AC(0.9)+think@2        acc 0.960 equiv   5578 ( +2024; 0.002538 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.9)+think@3        acc 0.954 equiv   5002 ( +1448; 0.002276 USD) | vs agree2 +1.4 [+1.1, +1.7]
+  AC(0.9)+think@4        acc 0.954 equiv   5592 ( +2038; 0.002544 USD) | vs agree2 +1.4 [+1.1, +1.6]
+  AC(0.95)               acc 0.876 equiv   7115 ( +3561; 0.003238 USD) | vs agree2 -6.5 [-7.8, -5.2]
+  AC(0.95)+think@2       acc 0.960 equiv   5578 ( +2024; 0.002538 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.95)+think@3       acc 0.960 equiv   6728 ( +3174; 0.003061 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.95)+think@4       acc 0.957 equiv   6319 ( +2765; 0.002875 USD) | vs agree2 +1.7 [+1.3, +2.1]
+  AC(0.99)               acc 0.876 equiv   8346 ( +4792; 0.003798 USD) | vs agree2 -6.4 [-7.8, -5.1]
+  AC(0.99)+think@2       acc 0.960 equiv   5578 ( +2024; 0.002538 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.99)+think@3       acc 0.960 equiv   6728 ( +3174; 0.003061 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  AC(0.99)+think@4       acc 0.960 equiv   7879 ( +4325; 0.003585 USD) | vs agree2 +2.0 [+1.3, +2.7]
+  always think           acc 0.960 equiv   3277 (  -277; 0.001491 USD) | vs agree2 +2.0 [+1.3, +2.7]
+R    4000: acc 0.847 equiv   1678
+R    8000: acc 0.910 equiv   2191
+R   16000: acc 0.940 equiv   2526
+R   24000: acc 0.945 equiv   2701
+R   32000: acc 0.954 equiv   2880
+R   48000: acc 0.957 equiv   3003
+R   64000: acc 0.958 equiv   3126
+R   96000: acc 0.959 equiv   3268
+parity: adaptive-SC Bayes at agree2's cost acc 0.958 equiv 3528 (-26) | vs agree2 +1.7 [+1.2, +2.3] | R [182332, 232323] | {'think:q32b_think@0': 3576, 'stop@2': 3251, 'stop@3': 1856, 'stop@4': 664, 'think:q32b_think@1': 649, 'stop@5': 282, 'think:q32b_think@3': 121, 'think:q32b_think@5': 116}
+split 1: adaptive-SC Bayes at agree2's cost acc 0.958 equiv 4444 (+890) | vs agree2 +1.8 [+1.1, +2.4] | R [377647, 256299] | {'think:q32b_think@0': 3984, 'stop@2': 3368, 'think:q32b_think@1': 1039, 'stop@4': 650, 'stop@3': 426, 'think:q32b_think@8': 358, 'think:q32b_think@2': 230, 'stop@5': 169}
+split 2: adaptive-SC Bayes at agree2's cost acc 0.959 equiv 3644 (+90) | vs agree2 +1.8 [+1.2, +2.4] | R [195464, 172168] | {'think:q32b_think@0': 3600, 'stop@3': 2999, 'stop@2': 1498, 'think:q32b_think@1': 1118, 'stop@4': 629, 'think:q32b_think@2': 268, 'think:q32b_think@3': 254, 'stop@5': 237}
+split 3: adaptive-SC Bayes at agree2's cost acc 0.957 equiv 3346 (-207) | vs agree2 +1.6 [+1.0, +2.3] | R [176829, 499998] | {'think:q32b_think@0': 5448, 'stop@2': 2492, 'stop@3': 1572, 'think:q32b_think@1': 870, 'stop@5': 206, 'think:q32b_think@2': 66, 'stop@7': 49, 'stop@4': 40}
+parity at agree2's accuracy: acc 0.941 cost 0.001174 USD vs 0.001617 | cost -27.4% [-30.1, -24.7] | acc +0.1 [-0.6, +0.7]
+split 1 at agree2's accuracy: acc 0.942 cost 0.001167 USD vs 0.001617 | cost -27.8% [-30.6, -25.2] | acc +0.1 [-0.5, +0.8]
+split 2 at agree2's accuracy: acc 0.942 cost 0.001166 USD vs 0.001617 | cost -27.9% [-30.6, -25.3] | acc +0.1 [-0.5, +0.8]
+split 3 at agree2's accuracy: acc 0.943 cost 0.001188 USD vs 0.001617 | cost -26.5% [-29.2, -24.0] | acc +0.2 [-0.4, +0.8]

analysis/bayes_sc_2026-10-25/v8/sc32_usd_openrouter_only.provenance.json

@@ -0,0 +1,92 @@
+{
+ "started_at": "2026-10-25T14:12:13+04:00",
+ "host": "ws006600",
+ "user": "vlad.smirnov",
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
+ "argv": [
+  "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
+  "--pool-dir",
+  "workdir/selector/pool_mth_k8_chat",
+  "--uhead",
+  "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+  "--task-prior",
+  "workdir/bayes/qprior/mth_qprior.json",
+  "--cost-model",
+  "usd",
+  "--price-scenario",
+  "openrouter_only",
+  "--tier",
+  "q32b_think=workdir/bayes/q32b_think_tier.json@qwen3-32b-fp8:think",
+  "--match-acc",
+  "--out",
+  "workdir/bayes/v8/sc32_usd_openrouter_only.json"
+ ],
+ "git": {
+  "commit": "f935c73993aae0c1137fdec2866aef2005007331",
+  "branch": "HEAD",
+  "subject": "Per-problem arrays in the Bayesian controllers; paired bootstrap against the hull mixture in currency_table",
+  "dirty": false,
+  "dirty_files": []
+ },
+ "gpus": {
+  "visible": "",
+  "devices": [
+   {
+    "index": "0",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
+   },
+   {
+    "index": "1",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
+   }
+  ]
+ },
+ "packages": {
+  "python": "3.11.15",
+  "torch": "2.9.1",
+  "vllm": "0.15.1",
+  "transformers": "4.57.6",
+  "speculators": "0.4.0.dev49",
+  "peft": null,
+  "accelerate": "1.13.0",
+  "numpy": "2.2.6",
+  "huggingface_hub": "0.36.2",
+  "openai": "2.32.0",
+  "wandb": "0.25.1"
+ },
+ "env": {
+  "CUDA_VISIBLE_DEVICES": ""
+ },
+ "models": [],
+ "files": {
+  "pool": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
+   "bytes": 65136887,
+   "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
+  },
+  "uhead": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+   "bytes": 9369632,
+   "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
+  },
+  "task_prior": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/qprior/mth_qprior.json",
+   "bytes": 28987,
+   "sha256": "897f405c643430dbe97023d9f8985e108330b82ed412a25ab2a4b84c992fbbd6"
+  },
+  "tier:q32b_think": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/q32b_think_tier.json",
+   "bytes": 26525,
+   "sha256": "4eef554b7fa4edff6cf11b63aadb44267d211bc25285eb260dcc1b45d2cf3a1f"
+  }
+ },
+ "finished_at": "2026-10-25T14:12:58+04:00",
+ "wall_secs": 44.9,
+ "n_problems": 1361
+}

analysis/bayes_sc_2026-10-25/v8/sc8_gpu_s.arrays.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5bcf76730b1e7d473f6a6d7ef77ff491829f527c97dc45db6956d7014bb8cc80
+size 327614

analysis/bayes_sc_2026-10-25/v8/sc8_gpu_s.json

@@ -0,0 +1,733 @@
+{
+ "ref": {
+  "acc": 0.929922850844967,
+  "equiv": 5447.452258972225
+ },
+ "cost_model": {
+  "currency": "gpu_s",
+  "unit": "GPU-seconds",
+  "per_equiv": 0.0006341154090044388,
+  "scenario": null,
+  "base": "qwen3-8b:nothink"
+ },
+ "baselines": {
+  "agree2": {
+   "acc": 0.929922850844967,
+   "equiv": 5447.452258972225,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "SC@1": {
+   "acc": 0.7907788390889052,
+   "equiv": 1149.6582624885161,
+   "delta_vs_agree2": [
+    -0.13914401175606173,
+    -0.15136044268919913,
+    -0.1273879500367377
+   ]
+  },
+  "SC@2": {
+   "acc": 0.8118111682586333,
+   "equiv": 2299.3165249770323,
+   "delta_vs_agree2": [
+    -0.11811168258633357,
+    -0.13019034717119765,
+    -0.10626262858192506
+   ]
+  },
+  "SC@3": {
+   "acc": 0.8444158706833211,
+   "equiv": 3448.9747874655477,
+   "delta_vs_agree2": [
+    -0.08550698016164585,
+    -0.09758908890521675,
+    -0.07425606171932403
+   ]
+  },
+  "SC@4": {
+   "acc": 0.8553453343130052,
+   "equiv": 4598.6330499540645,
+   "delta_vs_agree2": [
+    -0.07457751653196179,
+    -0.08610396767083027,
+    -0.06350799044819987
+   ]
+  },
+  "SC@5": {
+   "acc": 0.8630602498163116,
+   "equiv": 5748.29131244258,
+   "delta_vs_agree2": [
+    -0.0668626010286554,
+    -0.07866458486407053,
+    -0.055564612417340196
+   ]
+  },
+  "SC@6": {
+   "acc": 0.8693056576047025,
+   "equiv": 6897.949574931095,
+   "delta_vs_agree2": [
+    -0.06061719324026451,
+    -0.07214594048493755,
+    -0.04968428545187364
+   ]
+  },
+  "SC@7": {
+   "acc": 0.8717854518736223,
+   "equiv": 8047.607837419612,
+   "delta_vs_agree2": [
+    -0.0581373989713446,
+    -0.0704009000734754,
+    -0.04670279206465834
+   ]
+  },
+  "SC@8": {
+   "acc": 0.8761021307861866,
+   "equiv": 9197.266099908129,
+   "delta_vs_agree2": [
+    -0.05382072005878031,
+    -0.0664963721528288,
+    -0.041696133357825134
+   ]
+  },
+  "AC(0.8)": {
+   "acc": 0.8714180749448934,
+   "equiv": 4412.0116364692885,
+   "delta_vs_agree2": [
+    -0.058504775900073476,
+    -0.07026313372520206,
+    -0.04679463629684056
+   ]
+  },
+  "AC(0.8)+think@2": {
+   "acc": 0.929922850844967,
+   "equiv": 5447.452258972225,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@3": {
+   "acc": 0.9297391623806025,
+   "equiv": 5928.330676213297,
+   "delta_vs_agree2": [
+    -0.0001836884643644379,
+    -0.0005510653930933138,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@4": {
+   "acc": 0.9239529757531227,
+   "equiv": 5575.824379905499,
+   "delta_vs_agree2": [
+    -0.005969875091844232,
+    -0.008357825128581926,
+    -0.003857457751653196
+   ]
+  },
+  "AC(0.9)": {
+   "acc": 0.8757347538574578,
+   "equiv": 6165.396916034446,
+   "delta_vs_agree2": [
+    -0.054188096987509184,
+    -0.06658821638501101,
+    -0.04242973916238061
+   ]
+  },
+  "AC(0.9)+think@2": {
+   "acc": 0.9478324761204996,
+   "equiv": 10583.520622295184,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  },
+  "AC(0.9)+think@3": {
+   "acc": 0.9418626010286554,
+   "equiv": 7354.289774195097,
+   "delta_vs_agree2": [
+    0.011939750183688464,
+    0.009504729977957385,
+    0.014511388684790596
+   ]
+  },
+  "AC(0.9)+think@4": {
+   "acc": 0.9416789125642909,
+   "equiv": 7941.14717623332,
+   "delta_vs_agree2": [
+    0.011756061719324026,
+    0.009368111682586333,
+    0.014281778104335048
+   ]
+  },
+  "AC(0.95)": {
+   "acc": 0.8760102865540044,
+   "equiv": 7111.669323298678,
+   "delta_vs_agree2": [
+    -0.05391256429096253,
+    -0.06658706833210874,
+    -0.04192459588537841
+   ]
+  },
+  "AC(0.95)+think@2": {
+   "acc": 0.9478324761204996,
+   "equiv": 10583.520622295184,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  },
+  "AC(0.95)+think@3": {
+   "acc": 0.9478324761204996,
+   "equiv": 11733.178884783698,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  },
+  "AC(0.95)+think@4": {
+   "acc": 0.9447556943423953,
+   "equiv": 8926.147533635882,
+   "delta_vs_agree2": [
+    0.01483284349742836,
+    0.011572373254959589,
+    0.018369994489346064
+   ]
+  },
+  "AC(0.99)": {
+   "acc": 0.8761021307861866,
+   "equiv": 8341.71692481839,
+   "delta_vs_agree2": [
+    -0.05382072005878031,
+    -0.0664963721528288,
+    -0.041696133357825134
+   ]
+  },
+  "AC(0.99)+think@2": {
+   "acc": 0.9478324761204996,
+   "equiv": 10583.520622295184,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  },
+  "AC(0.99)+think@3": {
+   "acc": 0.9478324761204996,
+   "equiv": 11733.178884783698,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  },
+  "AC(0.99)+think@4": {
+   "acc": 0.9478324761204996,
+   "equiv": 12882.837147272214,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  },
+  "always think": {
+   "acc": 0.9478324761204996,
+   "equiv": 8284.204097318152,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  }
+ },
+ "frontier": [
+  {
+   "R": 4000,
+   "acc": 0.7907788390889052,
+   "equiv": 1150.6473977249664
+  },
+  {
+   "R": 8000,
+   "acc": 0.7907788390889052,
+   "equiv": 1150.6473977249664
+  },
+  {
+   "R": 16000,
+   "acc": 0.8124540778839089,
+   "equiv": 1506.2332346595122
+  },
+  {
+   "R": 24000,
+   "acc": 0.9054922850844966,
+   "equiv": 4730.2022522447505
+  },
+  {
+   "R": 32000,
+   "acc": 0.9287747979426892,
+   "equiv": 5469.786101770177
+  },
+  {
+   "R": 48000,
+   "acc": 0.9358927259368112,
+   "equiv": 5813.356175327136
+  },
+  {
+   "R": 64000,
+   "acc": 0.938372520205731,
+   "equiv": 5983.722530824361
+  },
+  {
+   "R": 96000,
+   "acc": 0.9397961058045555,
+   "equiv": 6185.602864908272
+  }
+ ],
+ "matched": [
+  {
+   "seed": 0,
+   "acc": 0.9066403379867745,
+   "equiv": 4780.387792227322,
+   "R": [
+    31737.957274763416,
+    25011.76896809135
+   ],
+   "delta_vs_agree2": [
+    -0.023282512858192505,
+    -0.02902392542248347,
+    -0.017496326230712712
+   ],
+   "actions": {
+    "stop@1": 5245,
+    "stop@2": 2352,
+    "stop@4": 145,
+    "think:think16k@0": 1864,
+    "stop@6": 42,
+    "stop@5": 179,
+    "think:think16k@5": 23,
+    "think:think16k@3": 632,
+    "think:think16k@6": 10,
+    "think:think16k@7": 6,
+    "stop@8": 18,
+    "stop@7": 34,
+    "stop@3": 319,
+    "think:think16k@4": 14,
+    "think:think16k@8": 5
+   }
+  },
+  {
+   "seed": 1,
+   "acc": 0.9310249816311535,
+   "equiv": 5613.162933157068,
+   "R": [
+    34211.01997690035,
+    30133.81637370327
+   ],
+   "delta_vs_agree2": [
+    0.0011021307861866275,
+    -0.004546289493019839,
+    0.006888317413666422
+   ],
+   "actions": {
+    "stop@2": 2909,
+    "stop@6": 34,
+    "stop@5": 118,
+    "stop@4": 289,
+    "think:think16k@0": 3688,
+    "stop@1": 3488,
+    "think:think16k@5": 72,
+    "stop@7": 66,
+    "think:think16k@2": 10,
+    "stop@3": 62,
+    "think:think16k@6": 48,
+    "think:think16k@8": 10,
+    "think:think16k@3": 70,
+    "stop@8": 20,
+    "think:think16k@4": 4
+   }
+  },
+  {
+   "seed": 2,
+   "acc": 0.9270756796473182,
+   "equiv": 5450.815983723787,
+   "R": [
+    34322.76200442399,
+    24175.51909619871
+   ],
+   "delta_vs_agree2": [
+    -0.002847171197648788,
+    -0.008496739529757531,
+    0.0030779298310066065
+   ],
+   "actions": {
+    "stop@5": 158,
+    "stop@4": 134,
+    "stop@2": 2611,
+    "stop@1": 3926,
+    "think:think16k@3": 117,
+    "think:think16k@2": 42,
+    "stop@7": 47,
+    "think:think16k@0": 3600,
+    "stop@3": 141,
+    "stop@6": 37,
+    "think:think16k@8": 6,
+    "think:think16k@7": 4,
+    "stop@8": 34,
+    "think:think16k@4": 13,
+    "think:think16k@6": 11,
+    "think:think16k@5": 7
+   }
+  },
+  {
+   "seed": 3,
+   "acc": 0.9276726671565025,
+   "equiv": 5433.844433047453,
+   "R": [
+    29929.841905114903,
+    24900.139139552528
+   ],
+   "delta_vs_agree2": [
+    -0.0022501836884643643,
+    -0.008175284717119764,
+    0.003950450036737686
+   ],
+   "actions": {
+    "stop@1": 3963,
+    "stop@2": 2547,
+    "think:think16k@3": 94,
+    "stop@4": 174,
+    "stop@5": 108,
+    "stop@6": 55,
+    "think:think16k@0": 3648,
+    "stop@3": 181,
+    "stop@7": 33,
+    "think:think16k@5": 37,
+    "stop@8": 18,
+    "think:think16k@6": 3,
+    "think:think16k@4": 18,
+    "think:think16k@8": 4,
+    "think:think16k@7": 5
+   }
+  }
+ ],
+ "matched_acc": [
+  {
+   "seed": 0,
+   "acc": 0.9296932402645114,
+   "equiv": 5526.845547149492,
+   "R": [
+    31738.04622725488,
+    33070.29237209457
+   ],
+   "delta_acc": [
+    -0.0002296105804555474,
+    -0.005923952975753123,
+    0.006016945260837613
+   ],
+   "rel_cost": [
+    0.014574388980922596,
+    -0.016265764902997956,
+    0.04722553277173239
+   ]
+  },
+  {
+   "seed": 1,
+   "acc": 0.9279481998530492,
+   "equiv": 5501.270490796819,
+   "R": [
+    24680.402832987907,
+    30133.801017737303
+   ],
+   "delta_acc": [
+    -0.0019746509919177074,
+    -0.007763133725202057,
+    0.004041146216017634
+   ],
+   "rel_cost": [
+    0.009879523356253932,
+    -0.022667434990197216,
+    0.04355080426891005
+   ]
+  },
+  {
+   "seed": 2,
+   "acc": 0.9262490815576782,
+   "equiv": 5406.323231147272,
+   "R": [
+    31828.864950143754,
+    24262.21860421506
+   ],
+   "delta_acc": [
+    -0.0036737692872887582,
+    -0.00936925973548861,
+    0.002344324026451126
+   ],
+   "rel_cost": [
+    -0.007550140114988803,
+    -0.039442974188090375,
+    0.025170127972970798
+   ]
+  },
+  {
+   "seed": 3,
+   "acc": 0.9276726671565025,
+   "equiv": 5433.844433047453,
+   "R": [
+    29929.86815727641,
+    24900.116367032453
+   ],
+   "delta_acc": [
+    -0.0022501836884643643,
+    -0.008175284717119764,
+    0.003950450036737686
+   ],
+   "rel_cost": [
+    -0.002498016554869073,
+    -0.03314613177154997,
+    0.029405375624232555
+   ]
+  }
+ ],
+ "transfer": {
+  "R": 30789.999548031326,
+  "m500": {
+   "acc": 0.932421875,
+   "equiv": 5548.312209962331,
+   "delta_vs_agree2": [
+    0.008203125,
+    -0.0109375,
+    0.0296875
+   ],
+   "prior_shift": null,
+   "delta": 0.0,
+   "baselines": {
+    "agree2": [
+     0.92421875,
+     5429.6649222993865
+    ],
+    "SC@1": [
+     0.78828125,
+     1129.850683748888
+    ],
+    "SC@2": [
+     0.81484375,
+     2259.701367497776
+    ],
+    "SC@3": [
+     0.84609375,
+     3389.5520512466637
+    ],
+    "SC@4": [
+     0.86171875,
+     4519.402734995552
+    ],
+    "SC@5": [
+     0.871875,
+     5649.25341874444
+    ],
+    "SC@6": [
+     0.8765625,
+     6779.104102493327
+    ],
+    "SC@7": [
+     0.88984375,
+     7908.954786242216
+    ],
+    "SC@8": [
+     0.89375,
+     9038.805469991104
+    ],
+    "AC(0.8)": [
+     0.88828125,
+     4191.5573709401015
+    ],
+    "AC(0.8)+think@2": [
+     0.92421875,
+     5429.6649222993865
+    ],
+    "AC(0.8)+think@3": [
+     0.92421875,
+     5913.907053378809
+    ],
+    "AC(0.8)+think@4": [
+     0.917578125,
+     5621.778754294677
+    ],
+    "AC(0.9)": [
+     0.89375,
+     6141.172027978451
+    ],
+    "AC(0.9)+think@2": [
+     0.9375,
+     10401.37537745888
+    ],
+    "AC(0.9)+think@3": [
+     0.9375,
+     7296.6699141805675
+    ],
+    "AC(0.9)+think@4": [
+     0.9375,
+     7894.798867542864
+    ],
+    "AC(0.95)": [
+     0.89375,
+     7059.039662375095
+    ],
+    "AC(0.95)+think@2": [
+     0.9375,
+     10401.37537745888
+    ],
+    "AC(0.95)+think@3": [
+     0.9375,
+     11531.226061207766
+    ],
+    "AC(0.95)+think@4": [
+     0.941796875,
+     8808.1838031147
+    ],
+    "AC(0.99)": [
+     0.89375,
+     8212.999457538364
+    ],
+    "AC(0.99)+think@2": [
+     0.9375,
+     10401.37537745888
+    ],
+    "AC(0.99)+think@3": [
+     0.9375,
+     11531.226061207766
+    ],
+    "AC(0.99)+think@4": [
+     0.9375,
+     12661.076744956656
+    ],
+    "always think": [
+     0.9375,
+     8141.6740099611025
+    ]
+   }
+  },
+  "aime": {
+   "acc": 0.24166666666666667,
+   "equiv": 3002.457782733884,
+   "delta_vs_agree2": [
+    -0.3125,
+    -0.40729166666666666,
+    -0.22083333333333333
+   ],
+   "prior_shift": null,
+   "delta": 0.0,
+   "baselines": {
+    "agree2": [
+     0.5541666666666667,
+     18915.787878948202
+    ],
+    "SC@1": [
+     0.22708333333333333,
+     2576.881539259918
+    ],
+    "SC@2": [
+     0.2625,
+     5153.763078519836
+    ],
+    "SC@3": [
+     0.28541666666666665,
+     7730.644617779754
+    ],
+    "SC@4": [
+     0.30625,
+     10307.526157039672
+    ],
+    "SC@5": [
+     0.3125,
+     12884.407696299588
+    ],
+    "SC@6": [
+     0.325,
+     15461.289235559509
+    ],
+    "SC@7": [
+     0.33958333333333335,
+     18038.17077481943
+    ],
+    "SC@8": [
+     0.35,
+     20615.052314079345
+    ],
+    "AC(0.8)": [
+     0.3458333333333333,
+     16312.521324942812
+    ],
+    "AC(0.8)+think@2": [
+     0.5541666666666667,
+     18915.787878948202
+    ],
+    "AC(0.8)+think@3": [
+     0.5541666666666667,
+     21018.56375089806
+    ],
+    "AC(0.8)+think@4": [
+     0.5354166666666667,
+     21935.201923610588
+    ],
+    "AC(0.9)": [
+     0.35,
+     19032.08558572155
+    ],
+    "AC(0.9)+think@2": [
+     0.5916666666666667,
+     22006.16625412832
+    ],
+    "AC(0.9)+think@3": [
+     0.5729166666666666,
+     22854.051555274233
+    ],
+    "AC(0.9)+think@4": [
+     0.5729166666666666,
+     25169.307578433563
+    ],
+    "AC(0.95)": [
+     0.35,
+     19899.130089981674
+    ],
+    "AC(0.95)+think@2": [
+     0.5916666666666667,
+     22006.16625412832
+    ],
+    "AC(0.95)+think@3": [
+     0.5916666666666667,
+     24583.047793388243
+    ],
+    "AC(0.95)+think@4": [
+     0.5822916666666667,
+     26054.21813973169
+    ],
+    "AC(0.99)": [
+     0.35,
+     20424.126233160696
+    ],
+    "AC(0.99)+think@2": [
+     0.5916666666666667,
+     22006.16625412832
+    ],
+    "AC(0.99)+think@3": [
+     0.5916666666666667,
+     24583.047793388243
+    ],
+    "AC(0.99)+think@4": [
+     0.5916666666666667,
+     27159.929332648164
+    ],
+    "always think": [
+     0.5916666666666667,
+     16852.403175608488
+    ]
+   }
+  }
+ }
+}

analysis/bayes_sc_2026-10-25/v8/sc8_gpu_s.log

@@ -0,0 +1,54 @@
+2026-10-25 14:12:15,520 online_rl_train INFO: ThinkBooster math evaluator loaded
+reference agree2 (rotation pairs, -> think16k): acc 0.930 equiv 5447
+cost model gpu_s: 1 equiv = 0.0006341 GPU-seconds; agree2 costs 3.454 GPU-seconds per problem
+  agree2                 acc 0.930 equiv   5447 (    +0; 3.454 GPU-seconds) | vs agree2 +0.0 [+0.0, +0.0]
+  SC@1                   acc 0.791 equiv   1150 ( -4298; 0.729 GPU-seconds) | vs agree2 -13.9 [-15.1, -12.7]
+  SC@2                   acc 0.812 equiv   2299 ( -3148; 1.458 GPU-seconds) | vs agree2 -11.8 [-13.0, -10.6]
+  SC@3                   acc 0.844 equiv   3449 ( -1998; 2.187 GPU-seconds) | vs agree2 -8.6 [-9.8, -7.4]
+  SC@4                   acc 0.855 equiv   4599 (  -849; 2.916 GPU-seconds) | vs agree2 -7.5 [-8.6, -6.4]
+  SC@5                   acc 0.863 equiv   5748 (  +301; 3.645 GPU-seconds) | vs agree2 -6.7 [-7.9, -5.6]
+  SC@6                   acc 0.869 equiv   6898 ( +1450; 4.374 GPU-seconds) | vs agree2 -6.1 [-7.2, -5.0]
+  SC@7                   acc 0.872 equiv   8048 ( +2600; 5.103 GPU-seconds) | vs agree2 -5.8 [-7.0, -4.7]
+  SC@8                   acc 0.876 equiv   9197 ( +3750; 5.832 GPU-seconds) | vs agree2 -5.4 [-6.6, -4.2]
+  AC(0.8)                acc 0.871 equiv   4412 ( -1035; 2.798 GPU-seconds) | vs agree2 -5.9 [-7.0, -4.7]
+  AC(0.8)+think@2        acc 0.930 equiv   5447 (    +0; 3.454 GPU-seconds) | vs agree2 +0.0 [+0.0, +0.0]
+  AC(0.8)+think@3        acc 0.930 equiv   5928 (  +481; 3.759 GPU-seconds) | vs agree2 -0.0 [-0.1, +0.0]
+  AC(0.8)+think@4        acc 0.924 equiv   5576 (  +128; 3.536 GPU-seconds) | vs agree2 -0.6 [-0.8, -0.4]
+  AC(0.9)                acc 0.876 equiv   6165 (  +718; 3.91 GPU-seconds) | vs agree2 -5.4 [-6.7, -4.2]
+  AC(0.9)+think@2        acc 0.948 equiv  10584 ( +5136; 6.711 GPU-seconds) | vs agree2 +1.8 [+1.2, +2.4]
+  AC(0.9)+think@3        acc 0.942 equiv   7354 ( +1907; 4.663 GPU-seconds) | vs agree2 +1.2 [+1.0, +1.5]
+  AC(0.9)+think@4        acc 0.942 equiv   7941 ( +2494; 5.036 GPU-seconds) | vs agree2 +1.2 [+0.9, +1.4]
+  AC(0.95)               acc 0.876 equiv   7112 ( +1664; 4.51 GPU-seconds) | vs agree2 -5.4 [-6.7, -4.2]
+  AC(0.95)+think@2       acc 0.948 equiv  10584 ( +5136; 6.711 GPU-seconds) | vs agree2 +1.8 [+1.2, +2.4]
+  AC(0.95)+think@3       acc 0.948 equiv  11733 ( +6286; 7.44 GPU-seconds) | vs agree2 +1.8 [+1.2, +2.4]
+  AC(0.95)+think@4       acc 0.945 equiv   8926 ( +3479; 5.66 GPU-seconds) | vs agree2 +1.5 [+1.2, +1.8]
+  AC(0.99)               acc 0.876 equiv   8342 ( +2894; 5.29 GPU-seconds) | vs agree2 -5.4 [-6.6, -4.2]
+  AC(0.99)+think@2       acc 0.948 equiv  10584 ( +5136; 6.711 GPU-seconds) | vs agree2 +1.8 [+1.2, +2.4]
+  AC(0.99)+think@3       acc 0.948 equiv  11733 ( +6286; 7.44 GPU-seconds) | vs agree2 +1.8 [+1.2, +2.4]
+  AC(0.99)+think@4       acc 0.948 equiv  12883 ( +7435; 8.169 GPU-seconds) | vs agree2 +1.8 [+1.2, +2.4]
+  always think           acc 0.948 equiv   8284 ( +2837; 5.253 GPU-seconds) | vs agree2 +1.8 [+1.2, +2.4]
+R    4000: acc 0.791 equiv   1151
+R    8000: acc 0.791 equiv   1151
+R   16000: acc 0.812 equiv   1506
+R   24000: acc 0.905 equiv   4730
+R   32000: acc 0.929 equiv   5470
+R   48000: acc 0.936 equiv   5813
+R   64000: acc 0.938 equiv   5984
+R   96000: acc 0.940 equiv   6186
+parity: adaptive-SC Bayes at agree2's cost acc 0.907 equiv 4780 (-667) | vs agree2 -2.3 [-2.9, -1.7] | R [31738, 25012] | {'stop@1': 5245, 'stop@2': 2352, 'think:think16k@0': 1864, 'think:think16k@3': 632, 'stop@3': 319, 'stop@5': 179, 'stop@4': 145, 'stop@6': 42}
+split 1: adaptive-SC Bayes at agree2's cost acc 0.931 equiv 5613 (+166) | vs agree2 +0.1 [-0.5, +0.7] | R [34211, 30134] | {'think:think16k@0': 3688, 'stop@1': 3488, 'stop@2': 2909, 'stop@4': 289, 'stop@5': 118, 'think:think16k@5': 72, 'think:think16k@3': 70, 'stop@7': 66}
+split 2: adaptive-SC Bayes at agree2's cost acc 0.927 equiv 5451 (+3) | vs agree2 -0.3 [-0.8, +0.3] | R [34323, 24176] | {'stop@1': 3926, 'think:think16k@0': 3600, 'stop@2': 2611, 'stop@5': 158, 'stop@3': 141, 'stop@4': 134, 'think:think16k@3': 117, 'stop@7': 47}
+split 3: adaptive-SC Bayes at agree2's cost acc 0.928 equiv 5434 (-14) | vs agree2 -0.2 [-0.8, +0.4] | R [29930, 24900] | {'stop@1': 3963, 'think:think16k@0': 3648, 'stop@2': 2547, 'stop@3': 181, 'stop@4': 174, 'stop@5': 108, 'think:think16k@3': 94, 'stop@6': 55}
+parity at agree2's accuracy: acc 0.930 cost 3.505 GPU-seconds vs 3.454 | cost +1.5% [-1.6, +4.7] | acc -0.0 [-0.6, +0.6]
+split 1 at agree2's accuracy: acc 0.928 cost 3.488 GPU-seconds vs 3.454 | cost +1.0% [-2.3, +4.4] | acc -0.2 [-0.8, +0.4]
+split 2 at agree2's accuracy: acc 0.926 cost 3.428 GPU-seconds vs 3.454 | cost -0.8% [-3.9, +2.5] | acc -0.4 [-0.9, +0.2]
+split 3 at agree2's accuracy: acc 0.928 cost 3.446 GPU-seconds vs 3.454 | cost -0.2% [-3.3, +2.9] | acc -0.2 [-0.8, +0.4]
+2026-10-25 14:13:10,256 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: '\\alpha^{85}-\\alpha^{-85}' vs '2\\cos(85\\theta)'
+2026-10-25 14:13:20,523 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '-2'
+2026-10-25 14:13:30,558 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '2\\cos(85\\theta)'
+2026-10-25 14:13:40,591 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '123'
+2026-10-25 14:13:50,625 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '\\alpha^{85}-\\alpha^{-85}'
+2026-10-25 14:14:00,960 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: '2' vs 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}'
+2026-10-25 14:14:11,369 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'L_{85}' vs 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}'
+transfer m500: adaptive-SC Bayes acc 0.932 equiv 5548 | agree2 0.924 at 5430 | vs agree2 +0.8 [-1.1, +3.0] | always think 0.938 at 8142
+transfer aime: adaptive-SC Bayes acc 0.242 equiv 3002 | agree2 0.554 at 18916 | vs agree2 -31.2 [-40.7, -22.1] | always think 0.592 at 16852

analysis/bayes_sc_2026-10-25/v8/sc8_gpu_s.provenance.json

@@ -0,0 +1,93 @@
+{
+ "started_at": "2026-10-25T14:12:13+04:00",
+ "host": "ws006600",
+ "user": "vlad.smirnov",
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
+ "argv": [
+  "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
+  "--pool-dir",
+  "workdir/selector/pool_mth_k8_chat",
+  "--uhead",
+  "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+  "--task-prior",
+  "workdir/bayes/qprior/mth_qprior.json",
+  "--cost-model",
+  "gpu_s",
+  "--tier",
+  "think16k=workdir/layer_probe/think16k_tier.json@qwen3-8b:think",
+  "--match-acc",
+  "--transfer",
+  "m500=workdir/selector/pool_m500_k8_chat,workdir/selector/pool_m500_k8_chat/uhead_selector.json,workdir/bayes/m500_tier.json,workdir/bayes/qprior/m500_qprior.json",
+  "aime=workdir/selector/pool_aime_k8_chat,workdir/selector/pool_aime_k8_chat/uhead_selector.json,workdir/bayes/aime_tier.json,workdir/bayes/qprior/aime_qprior.json",
+  "--out",
+  "workdir/bayes/v8/sc8_gpu_s.json"
+ ],
+ "git": {
+  "commit": "f935c73993aae0c1137fdec2866aef2005007331",
+  "branch": "HEAD",
+  "subject": "Per-problem arrays in the Bayesian controllers; paired bootstrap against the hull mixture in currency_table",
+  "dirty": false,
+  "dirty_files": []
+ },
+ "gpus": {
+  "visible": "",
+  "devices": [
+   {
+    "index": "0",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
+   },
+   {
+    "index": "1",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
+   }
+  ]
+ },
+ "packages": {
+  "python": "3.11.15",
+  "torch": "2.9.1",
+  "vllm": "0.15.1",
+  "transformers": "4.57.6",
+  "speculators": "0.4.0.dev49",
+  "peft": null,
+  "accelerate": "1.13.0",
+  "numpy": "2.2.6",
+  "huggingface_hub": "0.36.2",
+  "openai": "2.32.0",
+  "wandb": "0.25.1"
+ },
+ "env": {
+  "CUDA_VISIBLE_DEVICES": ""
+ },
+ "models": [],
+ "files": {
+  "pool": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
+   "bytes": 65136887,
+   "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
+  },
+  "uhead": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+   "bytes": 9369632,
+   "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
+  },
+  "task_prior": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/qprior/mth_qprior.json",
+   "bytes": 28987,
+   "sha256": "897f405c643430dbe97023d9f8985e108330b82ed412a25ab2a4b84c992fbbd6"
+  },
+  "tier:think16k": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
+   "bytes": 26666,
+   "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
+  }
+ },
+ "finished_at": "2026-10-25T14:14:16+04:00",
+ "wall_secs": 123.4,
+ "n_problems": 1361
+}

analysis/bayes_sc_2026-10-25/v8/sc8_tokens.arrays.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2eba426219c106e6a68272c8201ef8b1fd38cbdcf101502d40f23dc61f030461
+size 187669

analysis/bayes_sc_2026-10-25/v8/sc8_tokens.json

@@ -0,0 +1,736 @@
+{
+ "ref": {
+  "acc": 0.929922850844967,
+  "equiv": 4659.625211241734
+ },
+ "cost_model": {
+  "currency": "tokens",
+  "unit": "base tokens",
+  "per_equiv": 1.0,
+  "scenario": null,
+  "base": "qwen3-8b:nothink"
+ },
+ "baselines": {
+  "agree2": {
+   "acc": 0.929922850844967,
+   "equiv": 4659.625211241734,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "SC@1": {
+   "acc": 0.7907788390889052,
+   "equiv": 1124.9408523144746,
+   "delta_vs_agree2": [
+    -0.13914401175606173,
+    -0.15136044268919913,
+    -0.1273879500367377
+   ]
+  },
+  "SC@2": {
+   "acc": 0.8118111682586333,
+   "equiv": 2249.881704628949,
+   "delta_vs_agree2": [
+    -0.11811168258633357,
+    -0.13019034717119765,
+    -0.10626262858192506
+   ]
+  },
+  "SC@3": {
+   "acc": 0.8444158706833211,
+   "equiv": 3374.822556943424,
+   "delta_vs_agree2": [
+    -0.08550698016164585,
+    -0.09758908890521675,
+    -0.07425606171932403
+   ]
+  },
+  "SC@4": {
+   "acc": 0.8553453343130052,
+   "equiv": 4499.763409257898,
+   "delta_vs_agree2": [
+    -0.07457751653196179,
+    -0.08610396767083027,
+    -0.06350799044819987
+   ]
+  },
+  "SC@5": {
+   "acc": 0.8630602498163116,
+   "equiv": 5624.704261572373,
+   "delta_vs_agree2": [
+    -0.0668626010286554,
+    -0.07866458486407053,
+    -0.055564612417340196
+   ]
+  },
+  "SC@6": {
+   "acc": 0.8693056576047025,
+   "equiv": 6749.645113886848,
+   "delta_vs_agree2": [
+    -0.06061719324026451,
+    -0.07214594048493755,
+    -0.04968428545187364
+   ]
+  },
+  "SC@7": {
+   "acc": 0.8717854518736223,
+   "equiv": 7874.585966201323,
+   "delta_vs_agree2": [
+    -0.0581373989713446,
+    -0.0704009000734754,
+    -0.04670279206465834
+   ]
+  },
+  "SC@8": {
+   "acc": 0.8761021307861866,
+   "equiv": 8999.526818515797,
+   "delta_vs_agree2": [
+    -0.05382072005878031,
+    -0.0664963721528288,
+    -0.041696133357825134
+   ]
+  },
+  "AC(0.8)": {
+   "acc": 0.8714180749448934,
+   "equiv": 4328.9513225569435,
+   "delta_vs_agree2": [
+    -0.058504775900073476,
+    -0.07026313372520206,
+    -0.04679463629684056
+   ]
+  },
+  "AC(0.8)+think@2": {
+   "acc": 0.929922850844967,
+   "equiv": 4659.625211241734,
+   "delta_vs_agree2": [
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@3": {
+   "acc": 0.9297391623806025,
+   "equiv": 5133.4060525349005,
+   "delta_vs_agree2": [
+    -0.0001836884643644379,
+    -0.0005510653930933138,
+    0.0
+   ]
+  },
+  "AC(0.8)+think@4": {
+   "acc": 0.9239529757531227,
+   "equiv": 4968.986232549595,
+   "delta_vs_agree2": [
+    -0.005969875091844232,
+    -0.008357825128581926,
+    -0.003857457751653196
+   ]
+  },
+  "AC(0.9)": {
+   "acc": 0.8757347538574578,
+   "equiv": 6046.053545187362,
+   "delta_vs_agree2": [
+    -0.054188096987509184,
+    -0.06658821638501101,
+    -0.04242973916238061
+   ]
+  },
+  "AC(0.9)+think@2": {
+   "acc": 0.9478324761204996,
+   "equiv": 8591.141807494489,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  },
+  "AC(0.9)+think@3": {
+   "acc": 0.9418626010286554,
+   "equiv": 6364.151193975019,
+   "delta_vs_agree2": [
+    0.011939750183688464,
+    0.009504729977957385,
+    0.014511388684790596
+   ]
+  },
+  "AC(0.9)+think@4": {
+   "acc": 0.9416789125642909,
+   "equiv": 6941.938739897135,
+   "delta_vs_agree2": [
+    0.011756061719324026,
+    0.009368111682586333,
+    0.014281778104335048
+   ]
+  },
+  "AC(0.95)": {
+   "acc": 0.8760102865540044,
+   "equiv": 6970.480804555474,
+   "delta_vs_agree2": [
+    -0.05391256429096253,
+    -0.06658706833210874,
+    -0.04192459588537841
+   ]
+  },
+  "AC(0.95)+think@2": {
+   "acc": 0.9478324761204996,
+   "equiv": 8591.141807494489,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  },
+  "AC(0.95)+think@3": {
+   "acc": 0.9478324761204996,
+   "equiv": 9716.082659808964,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  },
+  "AC(0.95)+think@4": {
+   "acc": 0.9447556943423953,
+   "equiv": 7812.266908523145,
+   "delta_vs_agree2": [
+    0.01483284349742836,
+    0.011572373254959589,
+    0.018369994489346064
+   ]
+  },
+  "AC(0.99)": {
+   "acc": 0.8761021307861866,
+   "equiv": 8168.442964731815,
+   "delta_vs_agree2": [
+    -0.05382072005878031,
+    -0.0664963721528288,
+    -0.041696133357825134
+   ]
+  },
+  "AC(0.99)+think@2": {
+   "acc": 0.9478324761204996,
+   "equiv": 8591.141807494489,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  },
+  "AC(0.99)+think@3": {
+   "acc": 0.9478324761204996,
+   "equiv": 9716.082659808964,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  },
+  "AC(0.99)+think@4": {
+   "acc": 0.9478324761204996,
+   "equiv": 10841.023512123438,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  },
+  "always think": {
+   "acc": 0.9478324761204996,
+   "equiv": 6341.26010286554,
+   "delta_vs_agree2": [
+    0.017909625275532696,
+    0.0124896675238795,
+    0.023925422483468038
+   ]
+  }
+ },
+ "frontier": [
+  {
+   "R": 4000,
+   "acc": 0.7907788390889052,
+   "equiv": 1125.9210139603233
+  },
+  {
+   "R": 8000,
+   "acc": 0.7908706833210875,
+   "equiv": 1126.979803453343
+  },
+  {
+   "R": 16000,
+   "acc": 0.9173861131520941,
+   "equiv": 4030.2277369581184
+  },
+  {
+   "R": 24000,
+   "acc": 0.9252387950036738,
+   "equiv": 4310.205464731815
+  },
+  {
+   "R": 32000,
+   "acc": 0.9332292432035268,
+   "equiv": 4503.716642174872
+  },
+  {
+   "R": 48000,
+   "acc": 0.936627479794269,
+   "equiv": 4716.755602498163
+  },
+  {
+   "R": 64000,
+   "acc": 0.9396583394562822,
+   "equiv": 4888.65907421014
+  },
+  {
+   "R": 96000,
+   "acc": 0.9403012490815577,
+   "equiv": 5089.314639970609
+  }
+ ],
+ "matched": [
+  {
+   "seed": 0,
+   "acc": 0.9354794268919912,
+   "equiv": 4636.240145113887,
+   "R": [
+    52046.352193855826,
+    35185.86932298981
+   ],
+   "delta_vs_agree2": [
+    0.005556576047024247,
+    0.0007347538574577516,
+    0.010976533798677437
+   ],
+   "actions": {
+    "stop@2": 4094,
+    "stop@4": 122,
+    "stop@5": 219,
+    "think:think16k@0": 3576,
+    "stop@6": 37,
+    "think:think16k@3": 206,
+    "think:think16k@8": 6,
+    "think:think16k@6": 12,
+    "think:think16k@4": 60,
+    "stop@8": 19,
+    "think:think16k@7": 7,
+    "think:think16k@2": 49,
+    "stop@7": 61,
+    "think:think16k@5": 41,
+    "stop@3": 145,
+    "stop@1": 2234
+   }
+  },
+  {
+   "seed": 1,
+   "acc": 0.9348824393828068,
+   "equiv": 4736.318405584129,
+   "R": [
+    50402.42040091058,
+    43849.97673482865
+   ],
+   "delta_vs_agree2": [
+    0.004959588537839824,
+    -4.707016899338715e-05,
+    0.01047024246877296
+   ],
+   "actions": {
+    "stop@8": 46,
+    "stop@7": 46,
+    "stop@6": 41,
+    "stop@4": 191,
+    "stop@3": 93,
+    "stop@2": 4097,
+    "think:think16k@0": 3688,
+    "stop@1": 2121,
+    "stop@5": 224,
+    "think:think16k@3": 131,
+    "think:think16k@4": 62,
+    "think:think16k@2": 10,
+    "think:think16k@5": 55,
+    "think:think16k@7": 23,
+    "think:think16k@6": 49,
+    "think:think16k@8": 11
+   }
+  },
+  {
+   "seed": 2,
+   "acc": 0.9344232182218957,
+   "equiv": 4668.602066495224,
+   "R": [
+    46243.34481780861,
+    41763.58748877824
+   ],
+   "delta_vs_agree2": [
+    0.004500367376928729,
+    -0.0005510653930933138,
+    0.009690714548126372
+   ],
+   "actions": {
+    "stop@5": 285,
+    "stop@4": 74,
+    "stop@7": 30,
+    "stop@6": 33,
+    "stop@3": 78,
+    "stop@2": 4047,
+    "stop@1": 2263,
+    "think:think16k@3": 154,
+    "think:think16k@2": 152,
+    "think:think16k@4": 49,
+    "think:think16k@0": 3600,
+    "think:think16k@8": 11,
+    "think:think16k@7": 9,
+    "stop@8": 62,
+    "think:think16k@5": 21,
+    "think:think16k@6": 20
+   }
+  },
+  {
+   "seed": 3,
+   "acc": 0.936214180749449,
+   "equiv": 4662.522951873622,
+   "R": [
+    43408.50571447576,
+    42267.05589810954
+   ],
+   "delta_vs_agree2": [
+    0.006291329904481999,
+    0.001331741366642175,
+    0.011710139603232916
+   ],
+   "actions": {
+    "stop@1": 2454,
+    "stop@2": 3669,
+    "stop@3": 279,
+    "stop@6": 47,
+    "think:think16k@3": 137,
+    "stop@5": 299,
+    "think:think16k@0": 3648,
+    "think:think16k@4": 59,
+    "stop@7": 29,
+    "think:think16k@5": 33,
+    "think:think16k@8": 4,
+    "think:think16k@2": 166,
+    "think:think16k@6": 20,
+    "stop@8": 10,
+    "stop@4": 28,
+    "think:think16k@7": 6
+   }
+  }
+ ],
+ "matched_acc": [
+  {
+   "seed": 0,
+   "acc": 0.9318975018368847,
+   "equiv": 4448.155271858926,
+   "R": [
+    29578.903705614226,
+    27504.41122046006
+   ],
+   "delta_acc": [
+    0.0019746509919177074,
+    -0.003123851947097722,
+    0.007668993387215283
+   ],
+   "rel_cost": [
+    -0.04538346536382752,
+    -0.07336249426131876,
+    -0.016401739346024714
+   ]
+  },
+  {
+   "seed": 1,
+   "acc": 0.9291880969875091,
+   "equiv": 4460.377470609847,
+   "R": [
+    23569.059348243198,
+    27805.181973248178
+   ],
+   "delta_acc": [
+    -0.0007347538574577516,
+    -0.0061994856722997795,
+    0.004913666421748714
+   ],
+   "rel_cost": [
+    -0.042760464972845136,
+    -0.07338813066920981,
+    -0.0125749532621609
+   ]
+  },
+  {
+   "seed": 2,
+   "acc": 0.9293717854518736,
+   "equiv": 4385.907714915504,
+   "R": [
+    30793.590557203992,
+    21673.722457353742
+   ],
+   "delta_acc": [
+    -0.0005510653930933138,
+    -0.0057414125642909625,
+    0.005052580822924314
+   ],
+   "rel_cost": [
+    -0.05874238461622716,
+    -0.08726100304692518,
+    -0.02974360534001899
+   ]
+  },
+  {
+   "seed": 3,
+   "acc": 0.9310249816311535,
+   "equiv": 4410.375477590008,
+   "R": [
+    28154.412857283965,
+    23611.696452476262
+   ],
+   "delta_acc": [
+    0.0011021307861866275,
+    -0.004178912564290963,
+    0.006797621234386474
+   ],
+   "rel_cost": [
+    -0.05349136944542039,
+    -0.08083547747240183,
+    -0.024509551231061723
+   ]
+  }
+ ],
+ "transfer": {
+  "R": 45414.32571178388,
+  "m500": {
+   "acc": 0.938671875,
+   "equiv": 4885.79328125,
+   "delta_vs_agree2": [
+    0.014453125,
+    -0.003515625,
+    0.03438476562499995
+   ],
+   "prior_shift": null,
+   "delta": 0.0,
+   "baselines": {
+    "agree2": [
+     0.92421875,
+     4634.017890625
+    ],
+    "SC@1": [
+     0.78828125,
+     1103.76328125
+    ],
+    "SC@2": [
+     0.81484375,
+     2207.5265625
+    ],
+    "SC@3": [
+     0.84609375,
+     3311.28984375
+    ],
+    "SC@4": [
+     0.86171875,
+     4415.053125
+    ],
+    "SC@5": [
+     0.871875,
+     5518.81640625
+    ],
+    "SC@6": [
+     0.8765625,
+     6622.5796875
+    ],
+    "SC@7": [
+     0.88984375,
+     7726.34296875
+    ],
+    "SC@8": [
+     0.89375,
+     8830.10625
+    ],
+    "AC(0.8)": [
+     0.88828125,
+     4099.459375
+    ],
+    "AC(0.8)+think@2": [
+     0.92421875,
+     4634.017890625
+    ],
+    "AC(0.8)+think@3": [
+     0.92421875,
+     5108.860078125001
+    ],
+    "AC(0.8)+think@4": [
+     0.917578125,
+     4986.9370312500005
+    ],
+    "AC(0.9)": [
+     0.89375,
+     6009.02890625
+    ],
+    "AC(0.9)+think@2": [
+     0.9375,
+     8439.737187499999
+    ],
+    "AC(0.9)+think@3": [
+     0.9375,
+     6302.046171874999
+    ],
+    "AC(0.9)+think@4": [
+     0.9375,
+     6888.578984375001
+    ],
+    "AC(0.95)": [
+     0.89375,
+     6905.77578125
+    ],
+    "AC(0.95)+think@2": [
+     0.9375,
+     8439.737187499999
+    ],
+    "AC(0.95)+think@3": [
+     0.9375,
+     9543.50046875
+    ],
+    "AC(0.95)+think@4": [
+     0.941796875,
+     7697.955703125001
+    ],
+    "AC(0.99)": [
+     0.89375,
+     8028.82734375
+    ],
+    "AC(0.99)+think@2": [
+     0.9375,
+     8439.737187499999
+    ],
+    "AC(0.99)+think@3": [
+     0.9375,
+     9543.50046875
+    ],
+    "AC(0.99)+think@4": [
+     0.9375,
+     10647.26375
+    ],
+    "always think": [
+     0.9375,
+     6232.210625000001
+    ]
+   }
+  },
+  "aime": {
+   "acc": 0.5333333333333333,
+   "equiv": 17110.724166666663,
+   "delta_vs_agree2": [
+    -0.020833333333333332,
+    -0.04169270833333333,
+    -0.004166666666666667
+   ],
+   "prior_shift": null,
+   "delta": 0.0,
+   "baselines": {
+    "agree2": [
+     0.5541666666666667,
+     15601.996666666668
+    ],
+    "SC@1": [
+     0.22708333333333333,
+     2533.9333333333334
+    ],
+    "SC@2": [
+     0.2625,
+     5067.866666666667
+    ],
+    "SC@3": [
+     0.28541666666666665,
+     7601.8
+    ],
+    "SC@4": [
+     0.30625,
+     10135.733333333334
+    ],
+    "SC@5": [
+     0.3125,
+     12669.666666666666
+    ],
+    "SC@6": [
+     0.325,
+     15203.6
+    ],
+    "SC@7": [
+     0.33958333333333335,
+     17737.533333333333
+    ],
+    "SC@8": [
+     0.35,
+     20271.466666666667
+    ],
+    "AC(0.8)": [
+     0.3458333333333333,
+     16049.204166666666
+    ],
+    "AC(0.8)+think@2": [
+     0.5541666666666667,
+     15601.996666666668
+    ],
+    "AC(0.8)+think@3": [
+     0.5541666666666667,
+     17670.754999999997
+    ],
+    "AC(0.8)+think@4": [
+     0.5354166666666667,
+     18834.57125
+    ],
+    "AC(0.9)": [
+     0.35,
+     18722.379166666666
+    ],
+    "AC(0.9)+think@2": [
+     0.5916666666666667,
+     17967.561666666665
+    ],
+    "AC(0.9)+think@3": [
+     0.5729166666666666,
+     19178.001458333332
+    ],
+    "AC(0.9)+think@4": [
+     0.5729166666666666,
+     21455.782708333332
+    ],
+    "AC(0.95)": [
+     0.35,
+     19571.391666666666
+    ],
+    "AC(0.95)+think@2": [
+     0.5916666666666667,
+     17967.561666666665
+    ],
+    "AC(0.95)+think@3": [
+     0.5916666666666667,
+     20501.495
+    ],
+    "AC(0.95)+think@4": [
+     0.5822916666666667,
+     22189.036666666667
+    ],
+    "AC(0.99)": [
+     0.35,
+     20084.541666666668
+    ],
+    "AC(0.99)+think@2": [
+     0.5916666666666667,
+     17967.561666666665
+    ],
+    "AC(0.99)+think@3": [
+     0.5916666666666667,
+     20501.495
+    ],
+    "AC(0.99)+think@4": [
+     0.5916666666666667,
+     23035.428333333333
+    ],
+    "always think": [
+     0.5916666666666667,
+     12899.695
+    ]
+   }
+  }
+ }
+}

analysis/bayes_sc_2026-10-25/v8/sc8_tokens.log

@@ -0,0 +1,54 @@
+2026-10-25 14:12:15,478 online_rl_train INFO: ThinkBooster math evaluator loaded
+reference agree2 (rotation pairs, -> think16k): acc 0.930 equiv 4660
+cost model tokens: 1 equiv = 1 base tokens; agree2 costs 4660 base tokens per problem
+  agree2                 acc 0.930 equiv   4660 (    +0; 4660 base tokens) | vs agree2 +0.0 [+0.0, +0.0]
+  SC@1                   acc 0.791 equiv   1125 ( -3535; 1125 base tokens) | vs agree2 -13.9 [-15.1, -12.7]
+  SC@2                   acc 0.812 equiv   2250 ( -2410; 2250 base tokens) | vs agree2 -11.8 [-13.0, -10.6]
+  SC@3                   acc 0.844 equiv   3375 ( -1285; 3375 base tokens) | vs agree2 -8.6 [-9.8, -7.4]
+  SC@4                   acc 0.855 equiv   4500 (  -160; 4500 base tokens) | vs agree2 -7.5 [-8.6, -6.4]
+  SC@5                   acc 0.863 equiv   5625 (  +965; 5625 base tokens) | vs agree2 -6.7 [-7.9, -5.6]
+  SC@6                   acc 0.869 equiv   6750 ( +2090; 6750 base tokens) | vs agree2 -6.1 [-7.2, -5.0]
+  SC@7                   acc 0.872 equiv   7875 ( +3215; 7875 base tokens) | vs agree2 -5.8 [-7.0, -4.7]
+  SC@8                   acc 0.876 equiv   9000 ( +4340; 9000 base tokens) | vs agree2 -5.4 [-6.6, -4.2]
+  AC(0.8)                acc 0.871 equiv   4329 (  -331; 4329 base tokens) | vs agree2 -5.9 [-7.0, -4.7]
+  AC(0.8)+think@2        acc 0.930 equiv   4660 (    +0; 4660 base tokens) | vs agree2 +0.0 [+0.0, +0.0]
+  AC(0.8)+think@3        acc 0.930 equiv   5133 (  +474; 5133 base tokens) | vs agree2 -0.0 [-0.1, +0.0]
+  AC(0.8)+think@4        acc 0.924 equiv   4969 (  +309; 4969 base tokens) | vs agree2 -0.6 [-0.8, -0.4]
+  AC(0.9)                acc 0.876 equiv   6046 ( +1386; 6046 base tokens) | vs agree2 -5.4 [-6.7, -4.2]
+  AC(0.9)+think@2        acc 0.948 equiv   8591 ( +3932; 8591 base tokens) | vs agree2 +1.8 [+1.2, +2.4]
+  AC(0.9)+think@3        acc 0.942 equiv   6364 ( +1705; 6364 base tokens) | vs agree2 +1.2 [+1.0, +1.5]
+  AC(0.9)+think@4        acc 0.942 equiv   6942 ( +2282; 6942 base tokens) | vs agree2 +1.2 [+0.9, +1.4]
+  AC(0.95)               acc 0.876 equiv   6970 ( +2311; 6970 base tokens) | vs agree2 -5.4 [-6.7, -4.2]
+  AC(0.95)+think@2       acc 0.948 equiv   8591 ( +3932; 8591 base tokens) | vs agree2 +1.8 [+1.2, +2.4]
+  AC(0.95)+think@3       acc 0.948 equiv   9716 ( +5056; 9716 base tokens) | vs agree2 +1.8 [+1.2, +2.4]
+  AC(0.95)+think@4       acc 0.945 equiv   7812 ( +3153; 7812 base tokens) | vs agree2 +1.5 [+1.2, +1.8]
+  AC(0.99)               acc 0.876 equiv   8168 ( +3509; 8168 base tokens) | vs agree2 -5.4 [-6.6, -4.2]
+  AC(0.99)+think@2       acc 0.948 equiv   8591 ( +3932; 8591 base tokens) | vs agree2 +1.8 [+1.2, +2.4]
+  AC(0.99)+think@3       acc 0.948 equiv   9716 ( +5056; 9716 base tokens) | vs agree2 +1.8 [+1.2, +2.4]
+  AC(0.99)+think@4       acc 0.948 equiv  10841 ( +6181; 1.084e+04 base tokens) | vs agree2 +1.8 [+1.2, +2.4]
+  always think           acc 0.948 equiv   6341 ( +1682; 6341 base tokens) | vs agree2 +1.8 [+1.2, +2.4]
+R    4000: acc 0.791 equiv   1126
+R    8000: acc 0.791 equiv   1127
+R   16000: acc 0.917 equiv   4030
+R   24000: acc 0.925 equiv   4310
+R   32000: acc 0.933 equiv   4504
+R   48000: acc 0.937 equiv   4717
+R   64000: acc 0.940 equiv   4889
+R   96000: acc 0.940 equiv   5089
+parity: adaptive-SC Bayes at agree2's cost acc 0.935 equiv 4636 (-23) | vs agree2 +0.6 [+0.1, +1.1] | R [52046, 35186] | {'stop@2': 4094, 'think:think16k@0': 3576, 'stop@1': 2234, 'stop@5': 219, 'think:think16k@3': 206, 'stop@3': 145, 'stop@4': 122, 'stop@7': 61}
+split 1: adaptive-SC Bayes at agree2's cost acc 0.935 equiv 4736 (+77) | vs agree2 +0.5 [-0.0, +1.0] | R [50402, 43850] | {'stop@2': 4097, 'think:think16k@0': 3688, 'stop@1': 2121, 'stop@5': 224, 'stop@4': 191, 'think:think16k@3': 131, 'stop@3': 93, 'think:think16k@4': 62}
+split 2: adaptive-SC Bayes at agree2's cost acc 0.934 equiv 4669 (+9) | vs agree2 +0.5 [-0.1, +1.0] | R [46243, 41764] | {'stop@2': 4047, 'think:think16k@0': 3600, 'stop@1': 2263, 'stop@5': 285, 'think:think16k@3': 154, 'think:think16k@2': 152, 'stop@3': 78, 'stop@4': 74}
+split 3: adaptive-SC Bayes at agree2's cost acc 0.936 equiv 4663 (+3) | vs agree2 +0.6 [+0.1, +1.2] | R [43409, 42267] | {'stop@2': 3669, 'think:think16k@0': 3648, 'stop@1': 2454, 'stop@5': 299, 'stop@3': 279, 'think:think16k@2': 166, 'think:think16k@3': 137, 'think:think16k@4': 59}
+parity at agree2's accuracy: acc 0.932 cost 4448 base tokens vs 4660 | cost -4.5% [-7.3, -1.6] | acc +0.2 [-0.3, +0.8]
+split 1 at agree2's accuracy: acc 0.929 cost 4460 base tokens vs 4660 | cost -4.3% [-7.3, -1.3] | acc -0.1 [-0.6, +0.5]
+split 2 at agree2's accuracy: acc 0.929 cost 4386 base tokens vs 4660 | cost -5.9% [-8.7, -3.0] | acc -0.1 [-0.6, +0.5]
+split 3 at agree2's accuracy: acc 0.931 cost 4410 base tokens vs 4660 | cost -5.3% [-8.1, -2.5] | acc +0.1 [-0.4, +0.7]
+2026-10-25 14:13:10,584 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: '\\alpha^{85}-\\alpha^{-85}' vs '2\\cos(85\\theta)'
+2026-10-25 14:13:20,848 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '-2'
+2026-10-25 14:13:30,886 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '2\\cos(85\\theta)'
+2026-10-25 14:13:40,922 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '123'
+2026-10-25 14:13:50,957 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}' vs '\\alpha^{85}-\\alpha^{-85}'
+2026-10-25 14:14:01,298 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: '2' vs 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}'
+2026-10-25 14:14:11,714 judge_selector WARNING: answer comparison timed out after 10s, counted unequal: 'L_{85}' vs 'z^{85}+\\frac{1}{z^{85}}=\\phi^{85}+(\\phi-1)^{85}'
+transfer m500: adaptive-SC Bayes acc 0.939 equiv 4886 | agree2 0.924 at 4634 | vs agree2 +1.4 [-0.4, +3.4] | always think 0.938 at 6232
+transfer aime: adaptive-SC Bayes acc 0.533 equiv 17111 | agree2 0.554 at 15602 | vs agree2 -2.1 [-4.2, -0.4] | always think 0.592 at 12900

analysis/bayes_sc_2026-10-25/v8/sc8_tokens.provenance.json

@@ -0,0 +1,93 @@
+{
+ "started_at": "2026-10-25T14:12:13+04:00",
+ "host": "ws006600",
+ "user": "vlad.smirnov",
+ "cwd": "/mnt/data/users/vlad.smirnov/rl-uhead-combos",
+ "argv": [
+  "/mnt/data/users/vlad.smirnov/rl-uhead-combos/rl_controller/scripts/bayes_sc.py",
+  "--pool-dir",
+  "workdir/selector/pool_mth_k8_chat",
+  "--uhead",
+  "workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+  "--task-prior",
+  "workdir/bayes/qprior/mth_qprior.json",
+  "--cost-model",
+  "tokens",
+  "--tier",
+  "think16k=workdir/layer_probe/think16k_tier.json",
+  "--match-acc",
+  "--transfer",
+  "m500=workdir/selector/pool_m500_k8_chat,workdir/selector/pool_m500_k8_chat/uhead_selector.json,workdir/bayes/m500_tier.json,workdir/bayes/qprior/m500_qprior.json",
+  "aime=workdir/selector/pool_aime_k8_chat,workdir/selector/pool_aime_k8_chat/uhead_selector.json,workdir/bayes/aime_tier.json,workdir/bayes/qprior/aime_qprior.json",
+  "--out",
+  "workdir/bayes/v8/sc8_tokens.json"
+ ],
+ "git": {
+  "commit": "f935c73993aae0c1137fdec2866aef2005007331",
+  "branch": "HEAD",
+  "subject": "Per-problem arrays in the Bayesian controllers; paired bootstrap against the hull mixture in currency_table",
+  "dirty": false,
+  "dirty_files": []
+ },
+ "gpus": {
+  "visible": "",
+  "devices": [
+   {
+    "index": "0",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-a15d9536-8962-7d07-d199-3fb785e80c83"
+   },
+   {
+    "index": "1",
+    "name": "NVIDIA RTX 5880 Ada Generation",
+    "memory": "49140 MiB",
+    "driver": "580.178.04",
+    "uuid": "GPU-168c29de-060d-b509-1ebd-6be5535c9d6a"
+   }
+  ]
+ },
+ "packages": {
+  "python": "3.11.15",
+  "torch": "2.9.1",
+  "vllm": "0.15.1",
+  "transformers": "4.57.6",
+  "speculators": "0.4.0.dev49",
+  "peft": null,
+  "accelerate": "1.13.0",
+  "numpy": "2.2.6",
+  "huggingface_hub": "0.36.2",
+  "openai": "2.32.0",
+  "wandb": "0.25.1"
+ },
+ "env": {
+  "CUDA_VISIBLE_DEVICES": ""
+ },
+ "models": [],
+ "files": {
+  "pool": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/pool_chains.jsonl",
+   "bytes": 65136887,
+   "sha256": "00f58ec1868cbe5ed10312cfa5649745c18e4a0bd121ee93bd40f0303ece4fda"
+  },
+  "uhead": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/selector/pool_mth_k8_chat/uhead_selector.json",
+   "bytes": 9369632,
+   "sha256": "9417d517828e3fc083b56dcb62dada5aeaec69d60684adafe2ce76807fd39e21"
+  },
+  "task_prior": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/bayes/qprior/mth_qprior.json",
+   "bytes": 28987,
+   "sha256": "897f405c643430dbe97023d9f8985e108330b82ed412a25ab2a4b84c992fbbd6"
+  },
+  "tier:think16k": {
+   "path": "/mnt/data/users/vlad.smirnov/rl-uhead-ood/workdir/layer_probe/think16k_tier.json",
+   "bytes": 26666,
+   "sha256": "fc3e1788b20be07c0b78bc3fad95f06ec7eb60258556a8f3c4373d16f23c557f"
+  }
+ },
+ "finished_at": "2026-10-25T14:14:17+04:00",
+ "wall_secs": 123.8,
+ "n_problems": 1361
+}