GPT-5.6-Sol three-tool hidden evaluation

Video_Games/results/gpt-5.6-sol-hidden-three-tools-v1/manifest.json

@@ -0,0 +1,95 @@
+{
+  "agent": "DiversityBundleSpecificAgent",
+  "category": "Video_Games",
+  "comp_embeddings_sha256": "d697ad96681c98ea385c88b5b816943f09eb6ec114d237f57e1a2814aa3df129",
+  "dataset": {
+    "frozen_evaluation_input_sha256": "478503003eaeaee81cd280e4045edb4da422dc520b59f986b45a6d6cc7ead13e",
+    "path": "Video_Games/agent_hidden.jsonl",
+    "query_count": 771,
+    "repo": "iaouali/amazon-benchmark"
+  },
+  "diversity": {
+    "agent": "DiversityBundleSpecificAgent",
+    "catalog_sha256": "e87bb1d026e96612a24422255c6236868b6fe5a6a34494ca719b0154c47a437b",
+    "diversity_policy": "prompt_requested_advisory_not_forced",
+    "instructions": "Diversity / near-duplicate check:\nBefore final selection, use get_substitute_products on your observed candidate shortlist,\nordered from highest to lowest priority, with similarity_threshold=0.95. Reserve a tool\nround for this check within the existing budget. It compares only products in the same\nsubcategory and returns kept items and removed near-duplicates. Treat this as advisory\nevidence: prefer kept alternatives for redundant roles, but do not sacrifice explicit\nuser requirements merely to increase variety. It does not establish compatibility or\nprove that products are interchangeable. If pruning leaves an uncovered role, select\nanother observed suitable candidate or search again if budget remains. Return exactly\nthe requested number of items. Never invent product IDs. The check is not forced by\nthe harness; actual tool use is recorded separately.",
+    "item_embeddings_sha256": "6c6217d0060837072426a5bbaf2a9c7347fcf0b9c246f4eb877c7b137740579d",
+    "output_schema": "three-tool extension; fixed threshold; no gold product IDs",
+    "partition": "agent_hidden",
+    "script_sha256": "5d7eb5b87004016c5a6d18790b44194b7ca880ca3cff97e60e7fe7db34bacef2",
+    "similarity_threshold": 0.95,
+    "subcategory_sha256": "f59c01aa325602ba35d0cab692ff182bba881f8d7c75a4e8f969620cbf49c953",
+    "tool_sha256": "0b2be5f07658f1099be6ab38ffad18aa17f2f325af31ae1d888226ebdd5b53a1",
+    "tools": [
+      "search_products",
+      "get_complementary_products",
+      "get_substitute_products"
+    ],
+    "version": "bundle-three-tools-v1"
+  },
+  "frozen_prompt_sha256": "e44b033c73f33e300f851986a46e3dc39ebfcee3f0aefc6d2ab58ba243014999",
+  "max_output_tokens": 16384,
+  "max_tool_rounds": 6,
+  "min_comp_calls": 1,
+  "min_distinct_searches": 2,
+  "model": "gpt-5.6-sol",
+  "model_backend": "hosted API",
+  "publication_interval_seconds": 1800,
+  "query_count": 771,
+  "query_sha256": "478503003eaeaee81cd280e4045edb4da422dc520b59f986b45a6d6cc7ead13e",
+  "query_timeout_seconds": 1800.0,
+  "reasoning_effort": "medium",
+  "report_parser_sha256": "f60e57715ed8076bb9163054092f9f8a0213753a7c29cbe93ec2c54654a1e86a",
+  "retrieval_checkpoint_sha256": "4124c948bfee2e50cbeac46a52840b6e96af37efff1293d0e5fcf38886d8c28f",
+  "source_sweep_manifest_sha256": "17df1e13406257428f81494678badc1fa23bf6f4e15fe887c718b660b93a0e72",
+  "strict_bundle_json": false,
+  "temperature": 1.0,
+  "tool_device": "local B200",
+  "tool_release": {
+    "artifacts": {
+      "embeddings/comp_embeddings_qwen.pt": {
+        "sha256": "d697ad96681c98ea385c88b5b816943f09eb6ec114d237f57e1a2814aa3df129",
+        "size": 88550285
+      },
+      "metrics/complementarity_history_qwen.json": {
+        "sha256": null,
+        "size": 15109
+      },
+      "metrics/complementarity_split.json": {
+        "sha256": null,
+        "size": 1414
+      },
+      "metrics/retrieval_history_qwen.json": {
+        "sha256": null,
+        "size": 10647
+      },
+      "models/comp_model_qwen.pt": {
+        "sha256": "23e8929fc62522c197ed1650c000723c7c332176b8d55fdccb0ed2b4486ab329",
+        "size": 15779517
+      },
+      "models/recsys_model_qwen.pt": {
+        "sha256": "4124c948bfee2e50cbeac46a52840b6e96af37efff1293d0e5fcf38886d8c28f",
+        "size": 242754611
+      },
+      "queries/benchmark_split_metadata.json": {
+        "sha256": null,
+        "size": 916
+      },
+      "training_recipe.json": {
+        "sha256": null,
+        "size": 4426
+      }
+    },
+    "benchmark_fingerprint": "b8a70cafddd570fab09df2a275e0670da6a4f4218ce6f69981a0841432a7ec2a",
+    "comp_best_epoch": 15,
+    "comp_best_ndcg_at_10": 0.06810968596450241,
+    "repo": "iaouali/amazon-tools",
+    "retrieval_best_validation_multi_positive_mrr": 0.2355278730392456,
+    "retrieval_epochs_finished": 5,
+    "retrieval_selected_epoch_by_recorded_metric": 5,
+    "revision": "9774c4f28e96d0c3eb4f416a82d4c691482bbdf3",
+    "run_prefix": "Video_Games/runs/shared-train-best-recipes-v1",
+    "selection": "internal shared_train validation only",
+    "status": "complete"
+  }
+}

Video_Games/results/gpt-5.6-sol-hidden-three-tools-v1/results.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f61f27bd17de546264aa58f40f3aafaac7021e0ef69c17f6b1b4cd7664a037ec
+size 20

Video_Games/results/gpt-5.6-sol-hidden-three-tools-v1/summary.json

@@ -0,0 +1,24 @@
+{
+  "metrics": {
+    "complete_n": 0,
+    "diversity_use_rate": null,
+    "exact_item_hits": 0,
+    "expected_n": 771,
+    "recorded_attempts": 0,
+    "set_hit_all_expected_errors_as_misses": 0.0,
+    "set_hit_at_k": null,
+    "set_hits": 0,
+    "unresolved_n": 771
+  },
+  "model": "gpt-5.6-sol",
+  "partition": "agent_hidden",
+  "reasoning_effort": "medium",
+  "scoring": "SetHit@k: at least one exact reference item, k=reference size (3-7)",
+  "status": "incomplete_resumable",
+  "tools": [
+    "retrieval",
+    "complementarity",
+    "diversity"
+  ],
+  "updated_at": "2026-09-23T10:35:23.803101+00:00"
+}