Inference Scaling Evaluation Results Batch Conversion

data/frontiermath/anthropic/claude-opus-4-20250514/9facea1c-c815-4d43-a733-863a0fb4f86d.json

@@ -0,0 +1,114 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/anthropic_claude-opus-4-20250514/1786163764.623998",
+    "evaluation_timestamp": "1775078458.0",
+    "retrieved_timestamp": "1786163764.623998",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-20250514",
+        "id": "anthropic/claude-opus-4-20250514",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "accuracy on frontiermath/S-adaptive for scorer verification_code",
+            "source_data": {
+                "dataset_name": "frontiermath",
+                "source_type": "hf_dataset",
+                "samples_number": 12,
+                "sample_ids": [
+                    "ALL3",
+                    "CWA2",
+                    "CWD31",
+                    "FMT1",
+                    "LTI1",
+                    "ONO3",
+                    "OVE2",
+                    "PLD1",
+                    "PTA1",
+                    "RAP1",
+                    "RSG1",
+                    "TIK2"
+                ],
+                "additional_details": {
+                    "shuffled": "False"
+                }
+            },
+            "evaluation_timestamp": "1775078458.0",
+            "metric_config": {
+                "evaluation_description": "accuracy",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.09999999999999999,
+                "uncertainty": {
+                    "standard_error": {
+                        "value": 0.06628679652796171
+                    },
+                    "num_samples": 120
+                },
+                "details": {
+                    "total_matched_trajectories": "60",
+                    "stopping_reason_count_repetition_guard": "60",
+                    "avg_total_tokens_target_model": "384926.40",
+                    "total_total_tokens_target_model": "23095584.0",
+                    "avg_total_tokens_other_models": "117.00",
+                    "total_total_tokens_other_models": "7020.0",
+                    "avg_total_tokens_all_models": "385043.40",
+                    "total_total_tokens_all_models": "23102604.0",
+                    "avg_turn_count": "22.18",
+                    "total_turn_count": "1331.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/anthropic/claude-opus-4-20250514/9facea1c-c815-4d43-a733-863a0fb4f86d_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "dfbdae44c9daf8dab3479989b1143d4e93d03532d466c97ce375c09d2a43536c",
+        "total_rows": 120
+    }
+}

data/frontiermath/anthropic/claude-opus-4-20250514/9facea1c-c815-4d43-a733-863a0fb4f86d_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dfbdae44c9daf8dab3479989b1143d4e93d03532d466c97ce375c09d2a43536c
+size 17729364

data/frontiermath/anthropic/claude-opus-4-20250514/d52f5c7d-d8c0-40ff-9911-4152f539e032.json

@@ -0,0 +1,36 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/anthropic_claude-opus-4-20250514/1786166353.3642",
+    "evaluation_timestamp": "1775078458.0",
+    "retrieved_timestamp": "1786166353.3642",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-20250514",
+        "id": "anthropic/claude-opus-4-20250514",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/anthropic/claude-opus-4-20250514/d52f5c7d-d8c0-40ff-9911-4152f539e032_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "0151ae034de4c199dca4711974ef4c4f240efe53c291fb9d4f3811da91e26762",
+        "total_rows": 30
+    }
+}

data/frontiermath/anthropic/claude-opus-4-20250514/d52f5c7d-d8c0-40ff-9911-4152f539e032_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0151ae034de4c199dca4711974ef4c4f240efe53c291fb9d4f3811da91e26762
+size 4805137

data/frontiermath/anthropic/claude-opus-4-5-20251101/c4e3b3de-1fda-4354-8660-10c24a2d0e72.json

@@ -0,0 +1,114 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/anthropic_claude-opus-4-5-20251101/1786163701.80803",
+    "evaluation_timestamp": "1776165072.0",
+    "retrieved_timestamp": "1786163701.80803",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-5-20251101",
+        "id": "anthropic/claude-opus-4-5-20251101",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "accuracy on frontiermath/S-adaptive for scorer verification_code",
+            "source_data": {
+                "dataset_name": "frontiermath",
+                "source_type": "hf_dataset",
+                "samples_number": 12,
+                "sample_ids": [
+                    "ALL3",
+                    "CWA2",
+                    "CWD31",
+                    "FMT1",
+                    "LTI1",
+                    "ONO3",
+                    "OVE2",
+                    "PLD1",
+                    "PTA1",
+                    "RAP1",
+                    "RSG1",
+                    "TIK2"
+                ],
+                "additional_details": {
+                    "shuffled": "False"
+                }
+            },
+            "evaluation_timestamp": "1776165072.0",
+            "metric_config": {
+                "evaluation_description": "accuracy",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.5499999999999999,
+                "uncertainty": {
+                    "standard_error": {
+                        "value": 0.11381803659589923
+                    },
+                    "num_samples": 120
+                },
+                "details": {
+                    "total_matched_trajectories": "60",
+                    "stopping_reason_count_repetition_guard": "60",
+                    "avg_total_tokens_target_model": "1816732.38",
+                    "total_total_tokens_target_model": "109003943.0",
+                    "avg_total_tokens_other_models": "117.00",
+                    "total_total_tokens_other_models": "7020.0",
+                    "avg_total_tokens_all_models": "1816849.38",
+                    "total_total_tokens_all_models": "109010963.0",
+                    "avg_turn_count": "39.90",
+                    "total_turn_count": "2394.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/anthropic/claude-opus-4-5-20251101/c4e3b3de-1fda-4354-8660-10c24a2d0e72_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "335dfaea81938d499c2e0fee537bf9de060e8b93f228a1784bd5f541f4191ed3",
+        "total_rows": 120
+    }
+}

data/frontiermath/anthropic/claude-opus-4-5-20251101/c4e3b3de-1fda-4354-8660-10c24a2d0e72_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:335dfaea81938d499c2e0fee537bf9de060e8b93f228a1784bd5f541f4191ed3
+size 33594749

data/frontiermath/anthropic/claude-opus-4-6/11e17ed6-0552-4a54-a2ca-5c9cf901b2cf.json

@@ -0,0 +1,109 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/anthropic_claude-opus-4-6/1786183596.523224",
+    "evaluation_timestamp": "1775065496.0",
+    "retrieved_timestamp": "1786183596.523224",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-6",
+        "id": "anthropic/claude-opus-4-6",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "accuracy on frontiermath for scorer verification_code",
+            "source_data": {
+                "dataset_name": "frontiermath",
+                "source_type": "hf_dataset",
+                "samples_number": 12,
+                "sample_ids": [
+                    "ALL3",
+                    "PTA1"
+                ],
+                "additional_details": {
+                    "shuffled": "False"
+                }
+            },
+            "evaluation_timestamp": "1775065496.0",
+            "metric_config": {
+                "evaluation_description": "accuracy",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.1,
+                "uncertainty": {
+                    "standard_error": {
+                        "value": 0.1
+                    },
+                    "num_samples": 20
+                },
+                "details": {
+                    "total_matched_trajectories": "10",
+                    "stopping_reason_count_token_limit": "10",
+                    "avg_total_tokens_target_model": "10043627.50",
+                    "total_total_tokens_target_model": "100436275.0",
+                    "avg_total_tokens_other_models": "0.00",
+                    "total_total_tokens_other_models": "0.0",
+                    "avg_total_tokens_all_models": "10043627.50",
+                    "total_total_tokens_all_models": "100436275.0",
+                    "avg_turn_count": "130.10",
+                    "total_turn_count": "1301.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "max_retries": "3",
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {},
+                    "max_attempts": 3
+                },
+                "additional_details": {
+                    "max_retries": "3"
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/anthropic/claude-opus-4-6/11e17ed6-0552-4a54-a2ca-5c9cf901b2cf_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "3bbc248ccd45867d2ad0c1607a4824636cfa003c8a11200768e1804c6801cfa6",
+        "total_rows": 20
+    }
+}

data/frontiermath/anthropic/claude-opus-4-6/11e17ed6-0552-4a54-a2ca-5c9cf901b2cf_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bbc248ccd45867d2ad0c1607a4824636cfa003c8a11200768e1804c6801cfa6
+size 40134600

data/frontiermath/anthropic/claude-opus-4-6/2c01fa95-6a15-4034-9e36-d84a2b6acf82.json

@@ -0,0 +1,36 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/anthropic_claude-opus-4-6/1786166331.055622",
+    "evaluation_timestamp": "1775558526.0",
+    "retrieved_timestamp": "1786166331.055622",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-6",
+        "id": "anthropic/claude-opus-4-6",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/anthropic/claude-opus-4-6/2c01fa95-6a15-4034-9e36-d84a2b6acf82_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "45a1e50eb714b98128d62b2d51e652938085fb0446b0016f7ff5d2d74d0daad4",
+        "total_rows": 50
+    }
+}

data/frontiermath/anthropic/claude-opus-4-6/2c01fa95-6a15-4034-9e36-d84a2b6acf82_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45a1e50eb714b98128d62b2d51e652938085fb0446b0016f7ff5d2d74d0daad4
+size 23060727

data/frontiermath/anthropic/claude-opus-4-6/57457cac-48ee-4eda-a92f-190b93308728.json

@@ -0,0 +1,36 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/anthropic_claude-opus-4-6/1786166345.209746",
+    "evaluation_timestamp": "1775078458.0",
+    "retrieved_timestamp": "1786166345.209746",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-6",
+        "id": "anthropic/claude-opus-4-6",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/anthropic/claude-opus-4-6/57457cac-48ee-4eda-a92f-190b93308728_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "6636d2324e4dd91ef6c5c7d0fd28becf609798695fb1a6a055be1ecf16612d08",
+        "total_rows": 20
+    }
+}

data/frontiermath/anthropic/claude-opus-4-6/57457cac-48ee-4eda-a92f-190b93308728_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6636d2324e4dd91ef6c5c7d0fd28becf609798695fb1a6a055be1ecf16612d08
+size 15234855

data/frontiermath/anthropic/claude-opus-4-6/69d29d86-6e05-4383-9b63-bcb29d746aa5.json

@@ -0,0 +1,111 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/anthropic_claude-opus-4-6/1786183562.236135",
+    "evaluation_timestamp": "1775116015.0",
+    "retrieved_timestamp": "1786183562.236135",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-6",
+        "id": "anthropic/claude-opus-4-6",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "accuracy on frontiermath for scorer verification_code",
+            "source_data": {
+                "dataset_name": "frontiermath",
+                "source_type": "hf_dataset",
+                "samples_number": 12,
+                "sample_ids": [
+                    "ALL3",
+                    "PTA1"
+                ],
+                "additional_details": {
+                    "shuffled": "False"
+                }
+            },
+            "evaluation_timestamp": "1775116015.0",
+            "metric_config": {
+                "evaluation_description": "accuracy",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.3,
+                "uncertainty": {
+                    "standard_error": {
+                        "value": 0.3
+                    },
+                    "num_samples": 20
+                },
+                "details": {
+                    "total_matched_trajectories": "10",
+                    "stopping_reason_count_token_limit": "5",
+                    "stopping_reason_count_completed_on_successful_submit": "4",
+                    "stopping_reason_count_repetition_guard": "1",
+                    "avg_total_tokens_target_model": "8205078.90",
+                    "total_total_tokens_target_model": "82050789.0",
+                    "avg_total_tokens_other_models": "11.70",
+                    "total_total_tokens_other_models": "117.0",
+                    "avg_total_tokens_all_models": "8205090.60",
+                    "total_total_tokens_all_models": "82050906.0",
+                    "avg_turn_count": "100.60",
+                    "total_turn_count": "1006.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "max_retries": "3",
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {},
+                    "max_attempts": 3
+                },
+                "additional_details": {
+                    "max_retries": "3"
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/anthropic/claude-opus-4-6/69d29d86-6e05-4383-9b63-bcb29d746aa5_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "baf683dcb54396ac46f8d215fcbd5dbe09b4deb54a2c29996b63c15ef3c1e57a",
+        "total_rows": 20
+    }
+}

data/frontiermath/anthropic/claude-opus-4-6/69d29d86-6e05-4383-9b63-bcb29d746aa5_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:baf683dcb54396ac46f8d215fcbd5dbe09b4deb54a2c29996b63c15ef3c1e57a
+size 39349218

data/frontiermath/anthropic/claude-opus-4-6/c767a1ac-d7b8-43f8-88a0-81f6a559d4c6.json

@@ -0,0 +1,115 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/anthropic_claude-opus-4-6/1786163747.490918",
+    "evaluation_timestamp": "1775078458.0",
+    "retrieved_timestamp": "1786163747.490918",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-6",
+        "id": "anthropic/claude-opus-4-6",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "accuracy on frontiermath/S-adaptive for scorer verification_code",
+            "source_data": {
+                "dataset_name": "frontiermath",
+                "source_type": "hf_dataset",
+                "samples_number": 12,
+                "sample_ids": [
+                    "ALL3",
+                    "CWA2",
+                    "CWD31",
+                    "FMT1",
+                    "LTI1",
+                    "ONO3",
+                    "OVE2",
+                    "PLD1",
+                    "PTA1",
+                    "RAP1",
+                    "RSG1",
+                    "TIK2"
+                ],
+                "additional_details": {
+                    "shuffled": "False"
+                }
+            },
+            "evaluation_timestamp": "1775078458.0",
+            "metric_config": {
+                "evaluation_description": "accuracy",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.8000000000000002,
+                "uncertainty": {
+                    "standard_error": {
+                        "value": 0.10224747162910902
+                    },
+                    "num_samples": 120
+                },
+                "details": {
+                    "total_matched_trajectories": "50",
+                    "stopping_reason_count_repetition_guard": "47",
+                    "stopping_reason_count_token_limit": "3",
+                    "avg_total_tokens_target_model": "2484297.78",
+                    "total_total_tokens_target_model": "124214889.0",
+                    "avg_total_tokens_other_models": "109.98",
+                    "total_total_tokens_other_models": "5499.0",
+                    "avg_total_tokens_all_models": "2484407.76",
+                    "total_total_tokens_all_models": "124220388.0",
+                    "avg_turn_count": "42.40",
+                    "total_turn_count": "2120.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/anthropic/claude-opus-4-6/c767a1ac-d7b8-43f8-88a0-81f6a559d4c6_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "160b1a5042112a1778ecdf231d296b36da213bbb4de46d3e1627ba6d2b4e4aef",
+        "total_rows": 120
+    }
+}

data/frontiermath/anthropic/claude-opus-4-6/c767a1ac-d7b8-43f8-88a0-81f6a559d4c6_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:160b1a5042112a1778ecdf231d296b36da213bbb4de46d3e1627ba6d2b4e4aef
+size 94648811

data/frontiermath/openai/gpt-5-2025-08-07/9515aa0d-d47b-434c-9d78-701b48467e16.json

@@ -0,0 +1,117 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/openai_gpt-5-2025-08-07/1786163689.077474",
+    "evaluation_timestamp": "1776165072.0",
+    "retrieved_timestamp": "1786163689.077474",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "openai/gpt-5-2025-08-07",
+        "id": "openai/gpt-5-2025-08-07",
+        "developer": "openai",
+        "inference_platform": "openai",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "accuracy on frontiermath/S-adaptive for scorer verification_code",
+            "source_data": {
+                "dataset_name": "frontiermath",
+                "source_type": "hf_dataset",
+                "samples_number": 12,
+                "sample_ids": [
+                    "ALL3",
+                    "CWA2",
+                    "CWD31",
+                    "FMT1",
+                    "LTI1",
+                    "ONO3",
+                    "OVE2",
+                    "PLD1",
+                    "PTA1",
+                    "RAP1",
+                    "RSG1",
+                    "TIK2"
+                ],
+                "additional_details": {
+                    "shuffled": "False"
+                }
+            },
+            "evaluation_timestamp": "1776165072.0",
+            "metric_config": {
+                "evaluation_description": "accuracy",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.6166666666666666,
+                "uncertainty": {
+                    "standard_error": {
+                        "value": 0.12722762548624686
+                    },
+                    "num_samples": 120
+                },
+                "details": {
+                    "total_matched_trajectories": "60",
+                    "stopping_reason_count_repetition_guard": "60",
+                    "avg_total_tokens_target_model": "544454.80",
+                    "total_total_tokens_target_model": "32667288.0",
+                    "avg_total_tokens_other_models": "117.00",
+                    "total_total_tokens_other_models": "7020.0",
+                    "avg_total_tokens_all_models": "544571.80",
+                    "total_total_tokens_all_models": "32674308.0",
+                    "avg_turn_count": "18.00",
+                    "total_turn_count": "1080.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": true,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"high\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                },
+                "additional_details": {
+                    "reasoning_effort": "\"high\""
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/openai/gpt-5-2025-08-07/9515aa0d-d47b-434c-9d78-701b48467e16_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "e0fea54d34e31009ab7a1448b3754c8ae92de09a23f89ea33f6a75e222d9f828",
+        "total_rows": 120
+    }
+}

data/frontiermath/openai/gpt-5-2025-08-07/9515aa0d-d47b-434c-9d78-701b48467e16_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0fea54d34e31009ab7a1448b3754c8ae92de09a23f89ea33f6a75e222d9f828
+size 30684092

data/frontiermath/openai/gpt-5.2-2025-12-11/3ab00b3f-e615-4357-9246-e0b29b2d12e7.json

@@ -0,0 +1,118 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/openai_gpt-5.2-2025-12-11/1786163673.367597",
+    "evaluation_timestamp": "1775078458.0",
+    "retrieved_timestamp": "1786163673.367597",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "openai/gpt-5.2-2025-12-11",
+        "id": "openai/gpt-5.2-2025-12-11",
+        "developer": "openai",
+        "inference_platform": "openai",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "accuracy on frontiermath/S-adaptive for scorer verification_code",
+            "source_data": {
+                "dataset_name": "frontiermath",
+                "source_type": "hf_dataset",
+                "samples_number": 12,
+                "sample_ids": [
+                    "ALL3",
+                    "CWA2",
+                    "CWD31",
+                    "FMT1",
+                    "LTI1",
+                    "ONO3",
+                    "OVE2",
+                    "PLD1",
+                    "PTA1",
+                    "RAP1",
+                    "RSG1",
+                    "TIK2"
+                ],
+                "additional_details": {
+                    "shuffled": "False"
+                }
+            },
+            "evaluation_timestamp": "1775078458.0",
+            "metric_config": {
+                "evaluation_description": "accuracy",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.7000000000000001,
+                "uncertainty": {
+                    "standard_error": {
+                        "value": 0.11214168527250507
+                    },
+                    "num_samples": 120
+                },
+                "details": {
+                    "total_matched_trajectories": "60",
+                    "stopping_reason_count_repetition_guard": "52",
+                    "stopping_reason_count_token_limit": "8",
+                    "avg_total_tokens_target_model": "2508969.57",
+                    "total_total_tokens_target_model": "150538174.0",
+                    "avg_total_tokens_other_models": "101.40",
+                    "total_total_tokens_other_models": "6084.0",
+                    "avg_total_tokens_all_models": "2509070.97",
+                    "total_total_tokens_all_models": "150544258.0",
+                    "avg_turn_count": "44.47",
+                    "total_turn_count": "2668.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": true,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"high\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                },
+                "additional_details": {
+                    "reasoning_effort": "\"high\""
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/openai/gpt-5.2-2025-12-11/3ab00b3f-e615-4357-9246-e0b29b2d12e7_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "693decafa542900a874910cdd8c14922dc45b92ec8a7ff3ee3feabe12535408a",
+        "total_rows": 120
+    }
+}

data/frontiermath/openai/gpt-5.2-2025-12-11/3ab00b3f-e615-4357-9246-e0b29b2d12e7_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:693decafa542900a874910cdd8c14922dc45b92ec8a7ff3ee3feabe12535408a
+size 60747959

data/frontiermath/openai/gpt-5.4-2026-03-05/07a94f05-2534-4fec-9b39-67257b064492.json

@@ -0,0 +1,110 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/openai_gpt-5.4-2026-03-05/1786183537.362833",
+    "evaluation_timestamp": "1775051661.0",
+    "retrieved_timestamp": "1786183537.362833",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "openai/gpt-5.4-2026-03-05",
+        "id": "openai/gpt-5.4-2026-03-05",
+        "developer": "openai",
+        "inference_platform": "openai",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "accuracy on frontiermath for scorer verification_code",
+            "source_data": {
+                "dataset_name": "frontiermath",
+                "source_type": "hf_dataset",
+                "samples_number": 12,
+                "sample_ids": [
+                    "ALL3",
+                    "PTA1"
+                ],
+                "additional_details": {
+                    "shuffled": "False"
+                }
+            },
+            "evaluation_timestamp": "1775051661.0",
+            "metric_config": {
+                "evaluation_description": "accuracy",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.75,
+                "uncertainty": {
+                    "standard_error": {
+                        "value": 0.25
+                    },
+                    "num_samples": 20
+                },
+                "details": {
+                    "total_matched_trajectories": "10",
+                    "stopping_reason_count_completed_on_successful_submit": "7",
+                    "stopping_reason_count_repetition_guard": "3",
+                    "avg_total_tokens_target_model": "3367408.60",
+                    "total_total_tokens_target_model": "33674086.0",
+                    "avg_total_tokens_other_models": "35.10",
+                    "total_total_tokens_other_models": "351.0",
+                    "avg_total_tokens_all_models": "3367443.70",
+                    "total_total_tokens_all_models": "33674437.0",
+                    "avg_turn_count": "61.40",
+                    "total_turn_count": "614.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"to_float\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "max_retries": "3",
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"high\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {},
+                    "max_attempts": 3
+                },
+                "additional_details": {
+                    "max_retries": "3"
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/openai/gpt-5.4-2026-03-05/07a94f05-2534-4fec-9b39-67257b064492_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "23b3c5620c44c649d63b84cc3304b9a2e7a97305f176b006843258381629cd9f",
+        "total_rows": 20
+    }
+}

data/frontiermath/openai/gpt-5.4-2026-03-05/07a94f05-2534-4fec-9b39-67257b064492_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23b3c5620c44c649d63b84cc3304b9a2e7a97305f176b006843258381629cd9f
+size 13547806

data/frontiermath/openai/gpt-5.4-2026-03-05/8f265e55-5ca3-46fa-8669-03a1a5147b15.json

@@ -0,0 +1,109 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/openai_gpt-5.4-2026-03-05/1786183572.496821",
+    "evaluation_timestamp": "1775043378.0",
+    "retrieved_timestamp": "1786183572.496821",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "openai/gpt-5.4-2026-03-05",
+        "id": "openai/gpt-5.4-2026-03-05",
+        "developer": "openai",
+        "inference_platform": "openai",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "accuracy on frontiermath for scorer verification_code",
+            "source_data": {
+                "dataset_name": "frontiermath",
+                "source_type": "hf_dataset",
+                "samples_number": 12,
+                "sample_ids": [
+                    "ALL3",
+                    "PTA1"
+                ],
+                "additional_details": {
+                    "shuffled": "False"
+                }
+            },
+            "evaluation_timestamp": "1775043378.0",
+            "metric_config": {
+                "evaluation_description": "accuracy",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.35,
+                "uncertainty": {
+                    "standard_error": {
+                        "value": 0.35
+                    },
+                    "num_samples": 20
+                },
+                "details": {
+                    "total_matched_trajectories": "10",
+                    "stopping_reason_count_repetition_guard": "10",
+                    "avg_total_tokens_target_model": "1868378.70",
+                    "total_total_tokens_target_model": "18683787.0",
+                    "avg_total_tokens_other_models": "117.00",
+                    "total_total_tokens_other_models": "1170.0",
+                    "avg_total_tokens_all_models": "1868495.70",
+                    "total_total_tokens_all_models": "18684957.0",
+                    "avg_turn_count": "44.70",
+                    "total_turn_count": "447.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "max_retries": "3",
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"high\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {},
+                    "max_attempts": 3
+                },
+                "additional_details": {
+                    "max_retries": "3"
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/openai/gpt-5.4-2026-03-05/8f265e55-5ca3-46fa-8669-03a1a5147b15_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "59c8fd117e2cc61d294bf3921dc20bd0ecc464a42d2da1f10b451ceea6a26685",
+        "total_rows": 20
+    }
+}

data/frontiermath/openai/gpt-5.4-2026-03-05/8f265e55-5ca3-46fa-8669-03a1a5147b15_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59c8fd117e2cc61d294bf3921dc20bd0ecc464a42d2da1f10b451ceea6a26685
+size 7969186

data/frontiermath/openai/gpt-5.4-2026-03-05/922b6357-a72b-4447-8fa0-8e5a84bf7809.json

@@ -0,0 +1,36 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/openai_gpt-5.4-2026-03-05/1786168273.360609",
+    "evaluation_timestamp": "1775733996.0",
+    "retrieved_timestamp": "1786168273.360609",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "openai/gpt-5.4-2026-03-05",
+        "id": "openai/gpt-5.4-2026-03-05",
+        "developer": "openai",
+        "inference_platform": "openai",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/openai/gpt-5.4-2026-03-05/922b6357-a72b-4447-8fa0-8e5a84bf7809_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "7b23f4834f6b1ae5eca2aa1822095c05cbd52d0d65de5eba1f606a9029054d98",
+        "total_rows": 40
+    }
+}

data/frontiermath/openai/gpt-5.4-2026-03-05/922b6357-a72b-4447-8fa0-8e5a84bf7809_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b23f4834f6b1ae5eca2aa1822095c05cbd52d0d65de5eba1f606a9029054d98
+size 5029355

data/frontiermath/openai/gpt-5.4-2026-03-05/cc26d837-94a3-446b-b4f0-c56cf438822b.json

@@ -0,0 +1,115 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "frontiermath/openai_gpt-5.4-2026-03-05/1786163680.854002",
+    "evaluation_timestamp": "1776076028.0",
+    "retrieved_timestamp": "1786163680.854002",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "openai/gpt-5.4-2026-03-05",
+        "id": "openai/gpt-5.4-2026-03-05",
+        "developer": "openai",
+        "inference_platform": "openai",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "accuracy on frontiermath/S-adaptive/+6ep for scorer verification_code",
+            "source_data": {
+                "dataset_name": "frontiermath",
+                "source_type": "hf_dataset",
+                "samples_number": 10,
+                "sample_ids": [
+                    "CWA2",
+                    "CWD31",
+                    "FMT1",
+                    "LTI1",
+                    "ONO3",
+                    "OVE2",
+                    "PLD1",
+                    "RAP1",
+                    "RSG1",
+                    "TIK2"
+                ],
+                "additional_details": {
+                    "shuffled": "False"
+                }
+            },
+            "evaluation_timestamp": "1776076028.0",
+            "metric_config": {
+                "evaluation_description": "accuracy",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.9166666666666666,
+                "uncertainty": {
+                    "standard_error": {
+                        "value": 0.03726779962499649
+                    },
+                    "num_samples": 60
+                },
+                "details": {
+                    "total_matched_trajectories": "50",
+                    "stopping_reason_count_repetition_guard": "50",
+                    "avg_total_tokens_target_model": "446292.70",
+                    "total_total_tokens_target_model": "22314635.0",
+                    "avg_total_tokens_other_models": "117.00",
+                    "total_total_tokens_other_models": "5850.0",
+                    "avg_total_tokens_all_models": "446409.70",
+                    "total_total_tokens_all_models": "22320485.0",
+                    "avg_turn_count": "15.82",
+                    "total_turn_count": "791.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": true,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [{\"type\": \"tool\", \"name\": \"python\", \"params\": {\"timeout\": 180}}], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": [null, null, {\"type\": \"tool\", \"name\": \"inference_scaling_paper/submit_answer\", \"params\": {}}, false, \"\\n\\n\", false], \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"high\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                },
+                "additional_details": {
+                    "reasoning_effort": "\"high\""
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/frontiermath/openai/gpt-5.4-2026-03-05/cc26d837-94a3-446b-b4f0-c56cf438822b_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "a4f1fb2dd882af760adcea46153d9aa9dceb2c0f94c64c23af92b49c1eb70b9e",
+        "total_rows": 60
+    }
+}

data/frontiermath/openai/gpt-5.4-2026-03-05/cc26d837-94a3-446b-b4f0-c56cf438822b_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4f1fb2dd882af760adcea46153d9aa9dceb2c0f94c64c23af92b49c1eb70b9e
+size 8884657

data/healthbench/anthropic/claude-opus-4-20250514/1f639db4-9cf2-47e6-9ccc-57d250c63a6f.json

@@ -0,0 +1,36 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786166760.26322",
+    "evaluation_timestamp": "1775665083.0",
+    "retrieved_timestamp": "1786166760.26322",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-20250514",
+        "id": "anthropic/claude-opus-4-20250514",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/1f639db4-9cf2-47e6-9ccc-57d250c63a6f_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "87e433f6531dab85b148ce73d0fc3a6284ddd678f91bdcdf6b6a3c3c82e0f8ec",
+        "total_rows": 231
+    }
+}

data/healthbench/anthropic/claude-opus-4-20250514/1f639db4-9cf2-47e6-9ccc-57d250c63a6f_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87e433f6531dab85b148ce73d0fc3a6284ddd678f91bdcdf6b6a3c3c82e0f8ec
+size 15931458

data/healthbench/anthropic/claude-opus-4-20250514/2079508e-db7c-4388-a494-5cc56c9f1296.json

@@ -0,0 +1,36 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786168659.654294",
+    "evaluation_timestamp": "1775646127.0",
+    "retrieved_timestamp": "1786168659.654294",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-20250514",
+        "id": "anthropic/claude-opus-4-20250514",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/2079508e-db7c-4388-a494-5cc56c9f1296_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "b66c7c0abbd4848a1f1d7c4fe45d68336a8674a17d48e6a4aaa8bc3e01a3cc67",
+        "total_rows": 103
+    }
+}

data/healthbench/anthropic/claude-opus-4-20250514/2079508e-db7c-4388-a494-5cc56c9f1296_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b66c7c0abbd4848a1f1d7c4fe45d68336a8674a17d48e6a4aaa8bc3e01a3cc67
+size 6163011

data/healthbench/anthropic/claude-opus-4-20250514/25827f1a-0fcc-466c-a792-a3b79b8d7257.json

@@ -0,0 +1,36 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786166740.887247",
+    "evaluation_timestamp": "1775733995.0",
+    "retrieved_timestamp": "1786166740.887247",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-20250514",
+        "id": "anthropic/claude-opus-4-20250514",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/25827f1a-0fcc-466c-a792-a3b79b8d7257_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "b8d44e5e053f3c90fa4e80ba52d0ceb019d882fd4260092b18935918bf1be91d",
+        "total_rows": 80
+    }
+}

data/healthbench/anthropic/claude-opus-4-20250514/25827f1a-0fcc-466c-a792-a3b79b8d7257_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b8d44e5e053f3c90fa4e80ba52d0ceb019d882fd4260092b18935918bf1be91d
+size 5680176

data/healthbench/anthropic/claude-opus-4-20250514/2eff1faa-53c6-432f-97ee-d98f780fd116.json

@@ -0,0 +1,36 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786168676.326648",
+    "evaluation_timestamp": "1776259832.0",
+    "retrieved_timestamp": "1786168676.326648",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-20250514",
+        "id": "anthropic/claude-opus-4-20250514",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/2eff1faa-53c6-432f-97ee-d98f780fd116_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "4a75272343530073e98ae6bcffeb3ba7a0fa8a3ad70192524fa1128851bb0db3",
+        "total_rows": 300
+    }
+}

data/healthbench/anthropic/claude-opus-4-20250514/2eff1faa-53c6-432f-97ee-d98f780fd116_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a75272343530073e98ae6bcffeb3ba7a0fa8a3ad70192524fa1128851bb0db3
+size 18610256

data/healthbench/anthropic/claude-opus-4-20250514/32d55a94-b2ff-4378-990c-98e0f6b0fbc7.json

@@ -0,0 +1,320 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786164002.965401",
+    "evaluation_timestamp": "1776779722.0",
+    "retrieved_timestamp": "1786164002.965401",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-20250514",
+        "id": "anthropic/claude-opus-4-20250514",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+1ep/bcd0f6be for scorer _scorer",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "20",
+                    "sample_ids": "99b22b39-30d5-40dc-82d9-c1cd3c793a44,54349c88-e80a-4728-b181-305f97905d59,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541"
+                }
+            },
+            "evaluation_timestamp": "1776779722.0",
+            "metric_config": {
+                "evaluation_description": "bootstrap_score",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.16470137811168195,
+                "uncertainty": {
+                    "standard_deviation": 0.04611975735947024,
+                    "num_samples": 20
+                },
+                "details": {
+                    "total_matched_trajectories": "20",
+                    "stopping_reason_count_repetition_guard": "20",
+                    "avg_total_tokens_target_model": "10262.00",
+                    "total_total_tokens_target_model": "205240.0",
+                    "avg_total_tokens_other_models": "53217.20",
+                    "total_total_tokens_other_models": "1064344.0",
+                    "avg_total_tokens_all_models": "63479.20",
+                    "total_total_tokens_all_models": "1269584.0",
+                    "avg_turn_count": "27.25",
+                    "total_turn_count": "545.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": true,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "timeout": "5400",
+                            "attempt_timeout": "5400",
+                            "max_connections": "10",
+                            "reasoning_effort": "\"high\"",
+                            "reasoning_tokens": "16000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                },
+                "additional_details": {
+                    "timeout": "5400",
+                    "max_connections": "10",
+                    "reasoning_effort": "\"high\""
+                }
+            }
+        },
+        {
+            "evaluation_name": "std on healthbench/S-adaptive/+1ep/bcd0f6be for scorer _scorer",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "20",
+                    "sample_ids": "99b22b39-30d5-40dc-82d9-c1cd3c793a44,54349c88-e80a-4728-b181-305f97905d59,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541"
+                }
+            },
+            "evaluation_timestamp": "1776779722.0",
+            "metric_config": {
+                "evaluation_description": "std",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.04611975735947024,
+                "uncertainty": {
+                    "standard_deviation": 0.04611975735947024,
+                    "num_samples": 20
+                },
+                "details": {
+                    "total_matched_trajectories": "20",
+                    "stopping_reason_count_repetition_guard": "20",
+                    "avg_total_tokens_target_model": "10262.00",
+                    "total_total_tokens_target_model": "205240.0",
+                    "avg_total_tokens_other_models": "53217.20",
+                    "total_total_tokens_other_models": "1064344.0",
+                    "avg_total_tokens_all_models": "63479.20",
+                    "total_total_tokens_all_models": "1269584.0",
+                    "avg_turn_count": "27.25",
+                    "total_turn_count": "545.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": true,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "timeout": "5400",
+                            "attempt_timeout": "5400",
+                            "max_connections": "10",
+                            "reasoning_effort": "\"high\"",
+                            "reasoning_tokens": "16000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                },
+                "additional_details": {
+                    "timeout": "5400",
+                    "max_connections": "10",
+                    "reasoning_effort": "\"high\""
+                }
+            }
+        },
+        {
+            "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+1ep/bcd0f6be for scorer _scorer",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "20",
+                    "sample_ids": "99b22b39-30d5-40dc-82d9-c1cd3c793a44,54349c88-e80a-4728-b181-305f97905d59,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541"
+                }
+            },
+            "evaluation_timestamp": "1776779722.0",
+            "metric_config": {
+                "evaluation_description": "criteria_met_rate",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.29831932773109243,
+                "uncertainty": {
+                    "standard_deviation": 0.04611975735947024,
+                    "num_samples": 20
+                },
+                "details": {
+                    "total_matched_trajectories": "20",
+                    "stopping_reason_count_repetition_guard": "20",
+                    "avg_total_tokens_target_model": "10262.00",
+                    "total_total_tokens_target_model": "205240.0",
+                    "avg_total_tokens_other_models": "53217.20",
+                    "total_total_tokens_other_models": "1064344.0",
+                    "avg_total_tokens_all_models": "63479.20",
+                    "total_total_tokens_all_models": "1269584.0",
+                    "avg_turn_count": "27.25",
+                    "total_turn_count": "545.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": true,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "timeout": "5400",
+                            "attempt_timeout": "5400",
+                            "max_connections": "10",
+                            "reasoning_effort": "\"high\"",
+                            "reasoning_tokens": "16000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                },
+                "additional_details": {
+                    "timeout": "5400",
+                    "max_connections": "10",
+                    "reasoning_effort": "\"high\""
+                }
+            }
+        },
+        {
+            "evaluation_name": "mean on healthbench/S-adaptive/+1ep/bcd0f6be for scorer healthbench_score",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "20",
+                    "sample_ids": "99b22b39-30d5-40dc-82d9-c1cd3c793a44,54349c88-e80a-4728-b181-305f97905d59,c63a4306-4003-45a6-8be1-d669457473a9,d3cbb579-70a5-46e8-95f6-72cc9afa4ba3,3852d050-c009-4e33-b953-fd734a5b5e73,5dc23430-dfbb-4794-9f77-0d4dd6093033,67e76447-f16a-4a65-aba7-fc7166fd9b63,a783e7a3-cecc-4d15-b18c-3f2ddfc2dc6c,62e9d35f-153d-4744-b97c-9abfb2a90b1e,04b6e0c8-2835-415e-865d-aca9f1cd3c57,7a283823-00ee-489d-8dbc-356e2857526f,e89e00c9-29e9-420c-813d-e06ddc39d585,51433193-b0c3-4725-a3a8-9dd15e969b0c,48c89cd1-0ecf-4738-8fcc-392269585bd3,92ea478e-6bf8-4ffb-b88a-f57ccf5f2346,36c6e780-d4a8-47b2-8a36-90ee4b97607f,5fb6812f-9816-4d5d-b7e0-4ec3e9df5dc1,c29c95d1-5b5e-4f53-9ec5-83e1f72b5aa5,b81c059a-1137-47cf-bab8-ab5eb8ff34d5,886864a6-a0e4-468f-aaf4-26cf8dc5e541"
+                }
+            },
+            "evaluation_timestamp": "1776779722.0",
+            "metric_config": {
+                "evaluation_description": "mean",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.16470137811168195,
+                "uncertainty": {
+                    "num_samples": 20
+                },
+                "details": {
+                    "total_matched_trajectories": "20",
+                    "stopping_reason_count_repetition_guard": "20",
+                    "avg_total_tokens_target_model": "10262.00",
+                    "total_total_tokens_target_model": "205240.0",
+                    "avg_total_tokens_other_models": "53217.20",
+                    "total_total_tokens_other_models": "1064344.0",
+                    "avg_total_tokens_all_models": "63479.20",
+                    "total_total_tokens_all_models": "1269584.0",
+                    "avg_turn_count": "27.25",
+                    "total_turn_count": "545.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": true,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "timeout": "5400",
+                            "attempt_timeout": "5400",
+                            "max_connections": "10",
+                            "reasoning_effort": "\"high\"",
+                            "reasoning_tokens": "16000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                },
+                "additional_details": {
+                    "timeout": "5400",
+                    "max_connections": "10",
+                    "reasoning_effort": "\"high\""
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/32d55a94-b2ff-4378-990c-98e0f6b0fbc7_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "c6aaa01277b3edfb042a1b251112cf68337caf106ef37960a487baa936369a94",
+        "total_rows": 20
+    }
+}

data/healthbench/anthropic/claude-opus-4-20250514/32d55a94-b2ff-4378-990c-98e0f6b0fbc7_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6aaa01277b3edfb042a1b251112cf68337caf106ef37960a487baa936369a94
+size 1352558

data/healthbench/anthropic/claude-opus-4-20250514/33c5023a-011f-4d05-9cfb-209650d0a787.json

@@ -0,0 +1,316 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786164006.225737",
+    "evaluation_timestamp": "1776682987.0",
+    "retrieved_timestamp": "1786164006.225737",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-20250514",
+        "id": "anthropic/claude-opus-4-20250514",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+2ep/3523d247 for scorer _scorer",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "18",
+                    "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0"
+                }
+            },
+            "evaluation_timestamp": "1776682987.0",
+            "metric_config": {
+                "evaluation_description": "bootstrap_score",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.15747113635599683,
+                "uncertainty": {
+                    "standard_deviation": 0.06751224916506945,
+                    "num_samples": 36
+                },
+                "details": {
+                    "total_matched_trajectories": "36",
+                    "stopping_reason_count_repetition_guard": "36",
+                    "avg_total_tokens_target_model": "9225.19",
+                    "total_total_tokens_target_model": "332107.0",
+                    "avg_total_tokens_other_models": "61537.97",
+                    "total_total_tokens_other_models": "2215367.0",
+                    "avg_total_tokens_all_models": "70763.17",
+                    "total_total_tokens_all_models": "2547474.0",
+                    "avg_turn_count": "33.94",
+                    "total_turn_count": "1222.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "timeout": "3600",
+                            "attempt_timeout": "3600",
+                            "max_connections": "10",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "32000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                },
+                "additional_details": {
+                    "timeout": "3600",
+                    "max_connections": "10"
+                }
+            }
+        },
+        {
+            "evaluation_name": "std on healthbench/S-adaptive/+2ep/3523d247 for scorer _scorer",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "18",
+                    "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0"
+                }
+            },
+            "evaluation_timestamp": "1776682987.0",
+            "metric_config": {
+                "evaluation_description": "std",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.06751224916506945,
+                "uncertainty": {
+                    "standard_deviation": 0.06751224916506945,
+                    "num_samples": 36
+                },
+                "details": {
+                    "total_matched_trajectories": "36",
+                    "stopping_reason_count_repetition_guard": "36",
+                    "avg_total_tokens_target_model": "9225.19",
+                    "total_total_tokens_target_model": "332107.0",
+                    "avg_total_tokens_other_models": "61537.97",
+                    "total_total_tokens_other_models": "2215367.0",
+                    "avg_total_tokens_all_models": "70763.17",
+                    "total_total_tokens_all_models": "2547474.0",
+                    "avg_turn_count": "33.94",
+                    "total_turn_count": "1222.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "timeout": "3600",
+                            "attempt_timeout": "3600",
+                            "max_connections": "10",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "32000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                },
+                "additional_details": {
+                    "timeout": "3600",
+                    "max_connections": "10"
+                }
+            }
+        },
+        {
+            "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+2ep/3523d247 for scorer _scorer",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "18",
+                    "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0"
+                }
+            },
+            "evaluation_timestamp": "1776682987.0",
+            "metric_config": {
+                "evaluation_description": "criteria_met_rate",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.3104166666666667,
+                "uncertainty": {
+                    "standard_deviation": 0.06751224916506945,
+                    "num_samples": 36
+                },
+                "details": {
+                    "total_matched_trajectories": "36",
+                    "stopping_reason_count_repetition_guard": "36",
+                    "avg_total_tokens_target_model": "9225.19",
+                    "total_total_tokens_target_model": "332107.0",
+                    "avg_total_tokens_other_models": "61537.97",
+                    "total_total_tokens_other_models": "2215367.0",
+                    "avg_total_tokens_all_models": "70763.17",
+                    "total_total_tokens_all_models": "2547474.0",
+                    "avg_turn_count": "33.94",
+                    "total_turn_count": "1222.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "timeout": "3600",
+                            "attempt_timeout": "3600",
+                            "max_connections": "10",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "32000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                },
+                "additional_details": {
+                    "timeout": "3600",
+                    "max_connections": "10"
+                }
+            }
+        },
+        {
+            "evaluation_name": "mean on healthbench/S-adaptive/+2ep/3523d247 for scorer healthbench_score",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "18",
+                    "sample_ids": "c3eeecfc-de96-480f-929d-ac4a5e5a541c,269074dc-9495-484b-b2f9-0f0943c0f816,b97b1bae-b09f-4073-8e16-540bba766c60,8e88f21a-24c4-4f9c-b848-faf5e10e31b9,a49eb9af-926e-4f7c-94d0-133729059116,1049130c-e9c9-461d-b080-90027bc011c0,13bb02e4-844e-4972-9200-2befa5f911d8,2694a753-2f57-41eb-a7f9-8c4c7acdd2e8,b160df9f-be98-4a4d-9cc8-11a612ffaf87,5e0a84d7-b972-484c-a649-469b15f0a90f,c7606b1b-ca4d-4a71-8277-bcddff09a653,20cdebb7-9189-4f09-ad21-bfc8e8b75680,f24935ac-0530-4063-b2f1-ac45cd265037,1b8f338a-2e98-4a42-ba80-cba266f4b9fa,6f827e79-05c4-4383-8ed4-16d36260f29f,c835bebe-e998-4dcd-843a-2de03d4aabc6,70afd1ac-85ea-4455-a5a2-15b35378e10a,aa7a760a-5173-4a39-af83-48d6ae2230f0"
+                }
+            },
+            "evaluation_timestamp": "1776682987.0",
+            "metric_config": {
+                "evaluation_description": "mean",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.15747113635599683,
+                "uncertainty": {
+                    "num_samples": 36
+                },
+                "details": {
+                    "total_matched_trajectories": "36",
+                    "stopping_reason_count_repetition_guard": "36",
+                    "avg_total_tokens_target_model": "9225.19",
+                    "total_total_tokens_target_model": "332107.0",
+                    "avg_total_tokens_other_models": "61537.97",
+                    "total_total_tokens_other_models": "2215367.0",
+                    "avg_total_tokens_all_models": "70763.17",
+                    "total_total_tokens_all_models": "2547474.0",
+                    "avg_turn_count": "33.94",
+                    "total_turn_count": "1222.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "timeout": "3600",
+                            "attempt_timeout": "3600",
+                            "max_connections": "10",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "32000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                },
+                "additional_details": {
+                    "timeout": "3600",
+                    "max_connections": "10"
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/33c5023a-011f-4d05-9cfb-209650d0a787_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "8431d5e360fab9b94fc335d237b50d5ae752aeb8d91aab0bdd5a10f92d06b827",
+        "total_rows": 36
+    }
+}

data/healthbench/anthropic/claude-opus-4-20250514/33c5023a-011f-4d05-9cfb-209650d0a787_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8431d5e360fab9b94fc335d237b50d5ae752aeb8d91aab0bdd5a10f92d06b827
+size 2648839

data/healthbench/anthropic/claude-opus-4-20250514/520232ba-d29b-4f34-a474-fe68aaf93a67.json

@@ -0,0 +1,36 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786168645.145171",
+    "evaluation_timestamp": "1775604415.0",
+    "retrieved_timestamp": "1786168645.145171",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.202,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-20250514",
+        "id": "anthropic/claude-opus-4-20250514",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/520232ba-d29b-4f34-a474-fe68aaf93a67_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "c05ac8ee4dd7efbd44a5d5a334945b30bf240694cda0c176b79bf41496078202",
+        "total_rows": 100
+    }
+}

data/healthbench/anthropic/claude-opus-4-20250514/520232ba-d29b-4f34-a474-fe68aaf93a67_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c05ac8ee4dd7efbd44a5d5a334945b30bf240694cda0c176b79bf41496078202
+size 6115815

data/healthbench/anthropic/claude-opus-4-20250514/54678cb0-5e81-4c48-955e-5f28d3e78b7e.json

@@ -0,0 +1,36 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786168651.403012",
+    "evaluation_timestamp": "1776076027.0",
+    "retrieved_timestamp": "1786168651.403012",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-20250514",
+        "id": "anthropic/claude-opus-4-20250514",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/54678cb0-5e81-4c48-955e-5f28d3e78b7e_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "2ab9c630c0386743b7fb4e161a39f706ad6f1814e8da1b33f0519962d517595a",
+        "total_rows": 20
+    }
+}

data/healthbench/anthropic/claude-opus-4-20250514/54678cb0-5e81-4c48-955e-5f28d3e78b7e_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ab9c630c0386743b7fb4e161a39f706ad6f1814e8da1b33f0519962d517595a
+size 1391035

data/healthbench/anthropic/claude-opus-4-20250514/79fef613-b36a-4bc5-8fff-32ad69d2250b.json

@@ -0,0 +1,292 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786164021.027369",
+    "evaluation_timestamp": "1776259833.0",
+    "retrieved_timestamp": "1786164021.027369",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-20250514",
+        "id": "anthropic/claude-opus-4-20250514",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+6ep/f1c68b87 for scorer _scorer",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "41",
+                    "sample_ids": "77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,9ab66439-8090-4f6c-ba74-874591dfd1a6,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e"
+                }
+            },
+            "evaluation_timestamp": "1776259833.0",
+            "metric_config": {
+                "evaluation_description": "bootstrap_score",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.2478166311997635,
+                "uncertainty": {
+                    "standard_deviation": 0.03141629629739583,
+                    "num_samples": 246
+                },
+                "details": {
+                    "total_matched_trajectories": "35",
+                    "stopping_reason_count_repetition_guard": "35",
+                    "avg_total_tokens_target_model": "9195.49",
+                    "total_total_tokens_target_model": "321842.0",
+                    "avg_total_tokens_other_models": "58641.60",
+                    "total_total_tokens_other_models": "2052456.0",
+                    "avg_total_tokens_all_models": "67837.09",
+                    "total_total_tokens_all_models": "2374298.0",
+                    "avg_turn_count": "30.06",
+                    "total_turn_count": "1052.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                }
+            }
+        },
+        {
+            "evaluation_name": "std on healthbench/S-adaptive/+6ep/f1c68b87 for scorer _scorer",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "41",
+                    "sample_ids": "77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,9ab66439-8090-4f6c-ba74-874591dfd1a6,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e"
+                }
+            },
+            "evaluation_timestamp": "1776259833.0",
+            "metric_config": {
+                "evaluation_description": "std",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.03141629629739583,
+                "uncertainty": {
+                    "standard_deviation": 0.03141629629739583,
+                    "num_samples": 246
+                },
+                "details": {
+                    "total_matched_trajectories": "35",
+                    "stopping_reason_count_repetition_guard": "35",
+                    "avg_total_tokens_target_model": "9195.49",
+                    "total_total_tokens_target_model": "321842.0",
+                    "avg_total_tokens_other_models": "58641.60",
+                    "total_total_tokens_other_models": "2052456.0",
+                    "avg_total_tokens_all_models": "67837.09",
+                    "total_total_tokens_all_models": "2374298.0",
+                    "avg_turn_count": "30.06",
+                    "total_turn_count": "1052.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                }
+            }
+        },
+        {
+            "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+6ep/f1c68b87 for scorer _scorer",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "41",
+                    "sample_ids": "77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,9ab66439-8090-4f6c-ba74-874591dfd1a6,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e"
+                }
+            },
+            "evaluation_timestamp": "1776259833.0",
+            "metric_config": {
+                "evaluation_description": "criteria_met_rate",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.3142946317103621,
+                "uncertainty": {
+                    "standard_deviation": 0.03141629629739583,
+                    "num_samples": 246
+                },
+                "details": {
+                    "total_matched_trajectories": "35",
+                    "stopping_reason_count_repetition_guard": "35",
+                    "avg_total_tokens_target_model": "9195.49",
+                    "total_total_tokens_target_model": "321842.0",
+                    "avg_total_tokens_other_models": "58641.60",
+                    "total_total_tokens_other_models": "2052456.0",
+                    "avg_total_tokens_all_models": "67837.09",
+                    "total_total_tokens_all_models": "2374298.0",
+                    "avg_turn_count": "30.06",
+                    "total_turn_count": "1052.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                }
+            }
+        },
+        {
+            "evaluation_name": "mean on healthbench/S-adaptive/+6ep/f1c68b87 for scorer healthbench_score",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "41",
+                    "sample_ids": "77dff7ec-2e02-4b93-86d6-5d7765b8dd3d,9ab66439-8090-4f6c-ba74-874591dfd1a6,762d9eca-4835-4447-a649-f91740c74bd0,5c1cf475-6fff-42e2-8508-7e1a14ea9f65,06fb659f-a4aa-4a97-a831-ed2cd2aa7e69,608770a0-440d-4349-9a1c-863e9f4d3e24,a58987e9-8b59-461a-a68b-20efa1d37d51,2d878243-5d98-489a-b8f2-8809f813b6cf,452a9534-b4fc-4483-94e0-d34b06b8e299,e690c779-3cb8-4271-8541-a699bb1bd475,d984a2a8-1209-42bb-8f59-001421c96f8f,5caa2d1f-3ab9-4e25-97a2-af29c2dbbcfb,7b1d2983-ab41-48f2-bdcd-454cbf3f81b8,b267896f-cfc0-4bbe-9fda-fdfa287d1f43,ff514fc4-bd34-4a3f-be3f-840b0e1f8091,a131d5e7-2a27-426f-90f0-25eae3c4b490,620b9bf1-f0f3-4d68-aa74-1560fb406a68,9b4c14e9-c404-4646-87b7-d3badae8a68b,9385d26b-bf44-4ff7-9410-e181413540dd,a57a4b5e-0f12-4103-8852-72bd2b7dba24,cf8040e2-39e3-4200-ab7f-b9b532310ac3,4fc7b71f-9591-4ae1-bfca-2929c7225182,da05b57a-d776-45b6-96af-1750e0254eec,ef446d15-d019-4c5a-a052-ae8cc3ea3b81,0fe4ea94-633d-4f73-985b-cbcd9f5a0270,cbb0ca6e-aa5e-4683-b107-d7dfa4cb60ce,ae6ec2ce-5c23-4e65-802f-cef9f1624d71,cd3dbc9f-a39c-4d13-af60-cba7dbc3baad,02423ab9-f3ed-4096-a7df-bfa3ebb40c7c,084b0c27-79c5-42f6-9cbd-690d0e79167a,42465540-8af1-423b-9b35-a31de1e1d49e,74a50705-7db2-44be-84dc-92b2d960d344,3b79284e-1143-46b4-b6a0-78e5448ff773,afbfc79d-6455-40b8-b475-1bcde4ef0cc9,6d705f80-8259-4ca1-8645-163f42c85a3c,129057dd-4bf4-472d-a8cf-a063dbed60ad,2833344d-bff9-4bbe-8fab-6b13d64c6b2a,48669cdc-1a64-4c99-a6e8-9a7181341fa2,be70739f-234a-4a0f-9b14-d15b8800577b,4047884e-4629-457c-9d77-130e937fae60,eb97bae4-430e-45cd-a065-2df3ab5c600e"
+                }
+            },
+            "evaluation_timestamp": "1776259833.0",
+            "metric_config": {
+                "evaluation_description": "mean",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.2478166311997635,
+                "uncertainty": {
+                    "num_samples": 246
+                },
+                "details": {
+                    "total_matched_trajectories": "35",
+                    "stopping_reason_count_repetition_guard": "35",
+                    "avg_total_tokens_target_model": "9195.49",
+                    "total_total_tokens_target_model": "321842.0",
+                    "avg_total_tokens_other_models": "58641.60",
+                    "total_total_tokens_other_models": "2052456.0",
+                    "avg_total_tokens_all_models": "67837.09",
+                    "total_total_tokens_all_models": "2374298.0",
+                    "avg_turn_count": "30.06",
+                    "total_turn_count": "1052.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/79fef613-b36a-4bc5-8fff-32ad69d2250b_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "7447911938d044c71ea3c86923ce1ab16f3fa856ec747946c7535319a8a09203",
+        "total_rows": 246
+    }
+}

data/healthbench/anthropic/claude-opus-4-20250514/79fef613-b36a-4bc5-8fff-32ad69d2250b_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7447911938d044c71ea3c86923ce1ab16f3fa856ec747946c7535319a8a09203
+size 16235786

data/healthbench/anthropic/claude-opus-4-20250514/7a7fc16f-3232-4796-95e4-9311c88b650f.json

@@ -0,0 +1,292 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786164028.299496",
+    "evaluation_timestamp": "1776165243.0",
+    "retrieved_timestamp": "1786164028.299496",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-20250514",
+        "id": "anthropic/claude-opus-4-20250514",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [
+        {
+            "evaluation_name": "bootstrap_score on healthbench/S-adaptive/+7ep for scorer _scorer",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "20",
+                    "sample_ids": "19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175"
+                }
+            },
+            "evaluation_timestamp": "1776165243.0",
+            "metric_config": {
+                "evaluation_description": "bootstrap_score",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.24432381746789958,
+                "uncertainty": {
+                    "standard_deviation": 0.05666380138192282,
+                    "num_samples": 140
+                },
+                "details": {
+                    "total_matched_trajectories": "15",
+                    "stopping_reason_count_repetition_guard": "15",
+                    "avg_total_tokens_target_model": "10679.27",
+                    "total_total_tokens_target_model": "160189.0",
+                    "avg_total_tokens_other_models": "81013.20",
+                    "total_total_tokens_other_models": "1215198.0",
+                    "avg_total_tokens_all_models": "91692.47",
+                    "total_total_tokens_all_models": "1375387.0",
+                    "avg_turn_count": "50.67",
+                    "total_turn_count": "760.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                }
+            }
+        },
+        {
+            "evaluation_name": "std on healthbench/S-adaptive/+7ep for scorer _scorer",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "20",
+                    "sample_ids": "19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175"
+                }
+            },
+            "evaluation_timestamp": "1776165243.0",
+            "metric_config": {
+                "evaluation_description": "std",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.05666380138192282,
+                "uncertainty": {
+                    "standard_deviation": 0.05666380138192282,
+                    "num_samples": 140
+                },
+                "details": {
+                    "total_matched_trajectories": "15",
+                    "stopping_reason_count_repetition_guard": "15",
+                    "avg_total_tokens_target_model": "10679.27",
+                    "total_total_tokens_target_model": "160189.0",
+                    "avg_total_tokens_other_models": "81013.20",
+                    "total_total_tokens_other_models": "1215198.0",
+                    "avg_total_tokens_all_models": "91692.47",
+                    "total_total_tokens_all_models": "1375387.0",
+                    "avg_turn_count": "50.67",
+                    "total_turn_count": "760.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                }
+            }
+        },
+        {
+            "evaluation_name": "criteria_met_rate on healthbench/S-adaptive/+7ep for scorer _scorer",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "20",
+                    "sample_ids": "19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175"
+                }
+            },
+            "evaluation_timestamp": "1776165243.0",
+            "metric_config": {
+                "evaluation_description": "criteria_met_rate",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.32115171650055374,
+                "uncertainty": {
+                    "standard_deviation": 0.05666380138192282,
+                    "num_samples": 140
+                },
+                "details": {
+                    "total_matched_trajectories": "15",
+                    "stopping_reason_count_repetition_guard": "15",
+                    "avg_total_tokens_target_model": "10679.27",
+                    "total_total_tokens_target_model": "160189.0",
+                    "avg_total_tokens_other_models": "81013.20",
+                    "total_total_tokens_other_models": "1215198.0",
+                    "avg_total_tokens_all_models": "91692.47",
+                    "total_total_tokens_all_models": "1375387.0",
+                    "avg_turn_count": "50.67",
+                    "total_turn_count": "760.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                }
+            }
+        },
+        {
+            "evaluation_name": "mean on healthbench/S-adaptive/+7ep for scorer healthbench_score",
+            "source_data": {
+                "dataset_name": "healthbench",
+                "source_type": "other",
+                "additional_details": {
+                    "shuffled": "False",
+                    "inspect_dataset_location": "/home/ubuntu/.cache/inspect_evals/healthbench/full_c37d29dcb36dd46fd8b36341c84a8d06.jsonl",
+                    "inspect_dataset_name": "full_c37d29dcb36dd46fd8b36341c84a8d06",
+                    "samples_number": "20",
+                    "sample_ids": "19ec4833-86e9-4166-8b82-d1da09f31fd7,7ebc830a-8dbd-489b-9d61-4d8bacf0db8d,c971f9d1-5f6a-464e-b282-41c8f0d82f58,5a6e4a41-3ea6-4050-a971-93433fe34877,78510411-e03b-4f93-a369-c7e72d2e4908,bbd759b4-8e4d-4db7-bc36-f2e069396c1c,8f2a65de-dea7-48e8-8adb-6194eca26c08,6d5f483c-3e86-456d-bfd5-4e28de699ae5,f05491d8-d160-4b2b-bd5b-59d757585e39,85408180-238e-4c1b-92c1-55d8ce6c3717,ecd7453e-1dc1-4e52-bb95-67e8e5788573,0e073591-b3e7-4dc8-94d0-dc5aa93ca35c,7637b987-8fe2-49d0-8372-1ebf17284d70,5b294937-13e5-424e-8bb4-5d1904a2344a,c5bf7fc3-dfb0-4b47-9bbe-222e87b952e2,83cf8f2d-2857-4f01-a283-9595d8f4ae8e,7a4548e6-38b7-48dd-9088-d7d1219f7852,e0d1b955-1150-457e-96d7-91a9481999cd,24f9a6e7-b214-4011-94c4-6502f249a621,6a97773b-c33c-4f2e-9ac3-fdd53de0f175"
+                }
+            },
+            "evaluation_timestamp": "1776165243.0",
+            "metric_config": {
+                "evaluation_description": "mean",
+                "lower_is_better": false,
+                "score_type": "continuous",
+                "min_score": 0.0,
+                "max_score": 1.0
+            },
+            "score_details": {
+                "score": 0.24432381746789958,
+                "uncertainty": {
+                    "num_samples": 140
+                },
+                "details": {
+                    "total_matched_trajectories": "15",
+                    "stopping_reason_count_repetition_guard": "15",
+                    "avg_total_tokens_target_model": "10679.27",
+                    "total_total_tokens_target_model": "160189.0",
+                    "avg_total_tokens_other_models": "81013.20",
+                    "total_total_tokens_other_models": "1215198.0",
+                    "avg_total_tokens_all_models": "91692.47",
+                    "total_total_tokens_all_models": "1375387.0",
+                    "avg_turn_count": "50.67",
+                    "total_turn_count": "760.0"
+                }
+            },
+            "generation_config": {
+                "generation_args": {
+                    "reasoning": false,
+                    "agentic_eval_config": {
+                        "available_tools": []
+                    },
+                    "eval_plan": {
+                        "name": "plan",
+                        "steps": [
+                            "{\"solver\": \"react\", \"params\": {\"name\": null, \"description\": null, \"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"model\": null, \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"retry_refusals\": null, \"compaction\": \"CompactionSummary\", \"truncation\": \"disabled\", \"approval\": null}, \"params_passed\": {\"prompt\": [\"\", \"\\nYou are part of a multi-agent system designed to make agent coordination and execution easy. Agents uses two primary abstraction: **Agents** and **Handoffs**. An agent encompasses instructions and tools and can hand off a conversation to another agent when appropriate. Handoffs are achieved by calling a handoff function,generally named `transfer_to_<agent_name>`. Transfers between agents are handled seamlessly in the background; do not mention or draw attention to these transfers in your conversation with the user.\\n\", \"\\nYou are a helpful assistant attempting to submit the best possible answer. You have several tools available to help with finding the answer. You will see the result of tool calls right after sending the message. If you need to perform multiple actions, you can always send more messages with additional tool calls. Do some reasoning before your actions, describing what tool calls you are going to use and how they fit into your plan.\\n\", \"\\nWhen you have completed the task and have an answer, call the {submit}() tool to report it.\\n\"], \"tools\": [], \"attempts\": [999, \"wrapper\", \"_never_correct\"], \"submit\": null, \"on_continue\": \"_guard\", \"compaction\": \"CompactionSummary\"}}"
+                        ],
+                        "config": {
+                            "attempt_timeout": "1200",
+                            "reasoning_effort": "\"xhigh\"",
+                            "reasoning_tokens": "64000"
+                        }
+                    },
+                    "eval_limits": {
+                        "token_limit": 10000000
+                    },
+                    "sandbox": {}
+                }
+            }
+        }
+    ],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/7a7fc16f-3232-4796-95e4-9311c88b650f_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "d5792ef1ef4684e34e55fc1c40f629301d27fc9e04258a45b1fffb0a44089271",
+        "total_rows": 140
+    }
+}

data/healthbench/anthropic/claude-opus-4-20250514/7a7fc16f-3232-4796-95e4-9311c88b650f_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5792ef1ef4684e34e55fc1c40f629301d27fc9e04258a45b1fffb0a44089271
+size 9723427

data/healthbench/anthropic/claude-opus-4-20250514/911c0647-f046-4eb8-a5ff-b647d9fe9da1.json

@@ -0,0 +1,36 @@
+{
+    "schema_version": "0.3.0",
+    "evaluation_id": "healthbench/anthropic_claude-opus-4-20250514/1786168706.797071",
+    "evaluation_timestamp": "1776696116.0",
+    "retrieved_timestamp": "1786168706.797071",
+    "source_metadata": {
+        "source_name": "How Inference Compute Shapes Frontier LLM Evaluation",
+        "source_type": "evaluation_run",
+        "source_organization_name": "UK AI Security Initiative",
+        "evaluator_relationship": "third_party",
+        "source_organization_url": "https://www.aisi.gov.uk/",
+        "source_organization_logo_url": "https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/663bd707cb0214d8b72951b5_5a103bfcb506b52b4e099f3dc675c649_AISI%20Logo%20Colour%20Dark.svg"
+    },
+    "eval_library": {
+        "name": "inspect_ai",
+        "version": "inspect_ai:0.3.206.dev19+g6beb95d9,inference_scaling_paper:0.1.0"
+    },
+    "model_info": {
+        "name": "anthropic/claude-opus-4-20250514",
+        "id": "anthropic/claude-opus-4-20250514",
+        "developer": "anthropic",
+        "inference_platform": "anthropic",
+        "additional_details": {
+            "deployment_type": "externally_managed",
+            "model_availability": "closed_weights"
+        }
+    },
+    "evaluation_results": [],
+    "detailed_evaluation_results": {
+        "format": "jsonl",
+        "file_path": "data/healthbench/anthropic/claude-opus-4-20250514/911c0647-f046-4eb8-a5ff-b647d9fe9da1_samples.jsonl",
+        "hash_algorithm": "sha256",
+        "checksum": "c816e3e0ae0671322d5015549356a3f5756fbcf7899b0e4a9e6b56ca170f56b4",
+        "total_rows": 310
+    }
+}

data/healthbench/anthropic/claude-opus-4-20250514/911c0647-f046-4eb8-a5ff-b647d9fe9da1_samples.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c816e3e0ae0671322d5015549356a3f5756fbcf7899b0e4a9e6b56ca170f56b4
+size 19736626