{
  "configurations": {
    "01c1c65e291a5f5a": {
      "agent": "mini_swe_agent_v2",
      "endpoint": "https://dashscope.aliyuncs.com/compatible-mode/v1",
      "expected_provider": null,
      "global_concurrency": 5,
      "infrastructure_retries": 0,
      "max_logical_llm_calls": 300,
      "max_tokens": 65536,
      "model": "qwen3.5-flash",
      "model_profile_id": "dashscope-beijing-qwen3.5-flash-thinking-8192-cache-bounded",
      "reasoning_effort": null,
      "reasoning_max_tokens": 8192,
      "response_model": "qwen3.5-flash",
      "source_protocol": "cooperbench-four-arm-77-1.7-dashscope-qwen35-cache-bounded-1",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "floating alias; backend version unavailable"
    },
    "106e78241d44d994": {
      "agent": "mini_swe_agent_v2",
      "endpoint": "https://dashscope.aliyuncs.com/compatible-mode/v1",
      "expected_provider": null,
      "global_concurrency": 10,
      "infrastructure_retries": 1,
      "max_logical_llm_calls": 300,
      "max_tokens": 65536,
      "model": "qwen3.5-flash",
      "model_profile_id": "dashscope-beijing-qwen3.5-flash-thinking-8192-cache-bounded",
      "reasoning_effort": null,
      "reasoning_max_tokens": 8192,
      "response_model": "qwen3.5-flash",
      "source_protocol": "cooperbench-four-arm-77-1.7-dashscope-qwen35-cache-bounded-1",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "floating alias; backend version unavailable"
    },
    "2620e3b66aead1c2": {
      "agent": "mini_swe_agent_v2",
      "endpoint": "https://dashscope.aliyuncs.com/compatible-mode/v1",
      "expected_provider": null,
      "global_concurrency": 12,
      "infrastructure_retries": 3,
      "max_logical_llm_calls": 300,
      "max_tokens": 65536,
      "model": "qwen3.5-flash",
      "model_profile_id": "dashscope-beijing-qwen3.5-flash-thinking-8192-cache",
      "reasoning_effort": null,
      "reasoning_max_tokens": 8192,
      "response_model": "qwen3.5-flash",
      "source_protocol": "cooperbench-four-arm-1.13-r5-dashscope-qwen35-cache",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "floating alias; backend version unavailable"
    },
    "410a99f7395e143f": {
      "agent": "mini_swe_agent_v2",
      "catalog_canonical_slug": "deepseek/deepseek-v4-flash-20260423",
      "endpoint": "https://openrouter.ai/api/v1",
      "expected_provider": "GMICloud",
      "global_concurrency": 8,
      "infrastructure_retries": 0,
      "max_logical_llm_calls": 300,
      "max_tokens": 65536,
      "model": "deepseek/deepseek-v4-flash",
      "model_profile_id": "openrouter-deepseek-v4-flash-0423-gmicloud-fp8-thinking-high",
      "reasoning_effort": "high",
      "response_model": "deepseek/deepseek-v4-flash",
      "source_protocol": "cooperbench-four-arm-1.12-r5-openrouter-gmicloud-fp8-thinking-high",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "OpenRouter dated slug and fixed GMICloud FP8 route; immutable backend weights unverified"
    },
    "56c7cdea8073681a": {
      "agent": "mini_swe_agent_v2",
      "budget_exhaustion_workspace_recovery": true,
      "empty_submission_workspace_recovery": true,
      "endpoint": "https://dashscope.aliyuncs.com/compatible-mode/v1",
      "expected_provider": null,
      "global_concurrency": 6,
      "infrastructure_retries": 3,
      "max_logical_llm_calls": 300,
      "max_tokens": 65536,
      "model": "qwen3.5-flash",
      "model_profile_id": "dashscope-beijing-qwen3.5-flash-thinking-8192-cache-bounded-workspace-recovery",
      "reasoning_effort": null,
      "reasoning_max_tokens": 8192,
      "response_model": "qwen3.5-flash",
      "source_protocol": "cooperbench-four-arm-77-1.7-dashscope-qwen35-cache-bounded-1-workspace-recovery-1",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "floating alias; backend version unavailable"
    },
    "5da24dad69af9ac5": {
      "agent": "mini_swe_agent_v2",
      "catalog_canonical_slug": "deepseek/deepseek-v4-flash-20260423",
      "endpoint": "https://openrouter.ai/api/v1",
      "expected_provider": "GMICloud",
      "global_concurrency": 5,
      "infrastructure_retries": 0,
      "max_logical_llm_calls": 300,
      "max_tokens": 65536,
      "model": "deepseek/deepseek-v4-flash",
      "model_profile_id": "openrouter-deepseek-v4-flash-0423-gmicloud-fp8-thinking-high",
      "reasoning_effort": "high",
      "response_model": "deepseek/deepseek-v4-flash",
      "source_protocol": "cooperbench-three-arm-77-1.6-openrouter-gmicloud-fp8-thinking-high",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "OpenRouter dated slug and fixed GMICloud FP8 route; immutable backend weights unverified"
    },
    "730849cc0bbbaf9e": {
      "agent": "mini_swe_agent_v2",
      "endpoint": "https://dashscope.aliyuncs.com/compatible-mode/v1",
      "expected_provider": null,
      "global_concurrency": 2,
      "infrastructure_retries": 1,
      "max_logical_llm_calls": 300,
      "max_tokens": 65536,
      "model": "qwen3.5-flash",
      "model_profile_id": "dashscope-beijing-qwen3.5-flash-thinking-8192-cache-bounded",
      "reasoning_effort": null,
      "reasoning_max_tokens": 8192,
      "response_model": "qwen3.5-flash",
      "source_protocol": "cooperbench-four-arm-77-1.7-dashscope-qwen35-cache-bounded-1",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "floating alias; backend version unavailable"
    },
    "7f1aa28d8824cf73": {
      "agent": "mini_swe_agent_v2",
      "catalog_canonical_slug": "deepseek/deepseek-v4-flash-20260423",
      "endpoint": "https://openrouter.ai/api/v1",
      "expected_provider": "GMICloud",
      "global_concurrency": 5,
      "infrastructure_retries": 0,
      "max_logical_llm_calls": 300,
      "max_tokens": 65536,
      "model": "deepseek/deepseek-v4-flash",
      "model_profile_id": "openrouter-deepseek-v4-flash-0423-gmicloud-fp8-thinking-high",
      "reasoning_effort": "high",
      "response_model": "deepseek/deepseek-v4-flash",
      "source_protocol": "cooperbench-four-arm-1.12-r5-openrouter-gmicloud-fp8-thinking-high",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "OpenRouter dated slug and fixed GMICloud FP8 route; immutable backend weights unverified"
    },
    "825f55f8bb93fff6": {
      "agent": "mini_swe_agent_v2",
      "budget_exhaustion_workspace_recovery": true,
      "empty_submission_workspace_recovery": true,
      "endpoint": "https://api.apiyi.com/v1",
      "expected_provider": null,
      "global_concurrency": 16,
      "infrastructure_retries": 3,
      "max_logical_llm_calls": 300,
      "model": "gemini-3.1-flash-lite",
      "model_profile_id": "apiyi-gemini-3.1-flash-lite-thinking-medium",
      "reasoning_effort": "medium",
      "response_model": "gemini-3.1-flash-lite",
      "source_protocol": "cooperbench-three-arm-77-1.11-apiyi-gemini31-flash-lite-thinking-medium-v3",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "APIYI gemini-3.1-flash-lite stable alias; immutable upstream weights are not independently attested"
    },
    "98be4ab55601bb40": {
      "agent": "mini_swe_agent_v2",
      "catalog_canonical_slug": "deepseek/deepseek-v4-flash-20260423",
      "endpoint": "https://openrouter.ai/api/v1",
      "expected_provider": "GMICloud",
      "global_concurrency": 10,
      "infrastructure_retries": 1,
      "max_logical_llm_calls": 300,
      "max_tokens": 65536,
      "model": "deepseek/deepseek-v4-flash",
      "model_profile_id": "openrouter-deepseek-v4-flash-0423-gmicloud-fp8-thinking-high",
      "reasoning_effort": "high",
      "response_model": "deepseek/deepseek-v4-flash",
      "source_protocol": "cooperbench-four-arm-1.12-r5-openrouter-gmicloud-fp8-thinking-high",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "OpenRouter dated slug and fixed GMICloud FP8 route; immutable backend weights unverified"
    },
    "9e7f1bbb272dd39e": {
      "agent": "mini_swe_agent_v2",
      "catalog_canonical_slug": "deepseek/deepseek-v4-flash-20260423",
      "endpoint": "https://openrouter.ai/api/v1",
      "expected_provider": "GMICloud",
      "global_concurrency": 10,
      "infrastructure_retries": 1,
      "max_logical_llm_calls": 300,
      "max_tokens": 65536,
      "model": "deepseek/deepseek-v4-flash",
      "model_profile_id": "openrouter-deepseek-v4-flash-0423-gmicloud-fp8-thinking-high",
      "reasoning_effort": "high",
      "response_model": "deepseek/deepseek-v4-flash",
      "source_protocol": "cooperbench-three-arm-77-1.6-openrouter-gmicloud-fp8-thinking-high",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "OpenRouter dated slug and fixed GMICloud FP8 route; immutable backend weights unverified"
    },
    "b62eb336168a3a60": {
      "agent": "mini_swe_agent_v2",
      "budget_exhaustion_workspace_recovery": true,
      "empty_submission_workspace_recovery": true,
      "endpoint": "https://dashscope.aliyuncs.com/compatible-mode/v1",
      "expected_provider": null,
      "global_concurrency": 10,
      "infrastructure_retries": 3,
      "max_logical_llm_calls": 300,
      "max_tokens": 65536,
      "model": "qwen3.5-flash",
      "model_profile_id": "dashscope-beijing-qwen3.5-flash-thinking-8192-cache-workspace-recovery",
      "reasoning_effort": null,
      "reasoning_max_tokens": 8192,
      "response_model": "qwen3.5-flash",
      "source_protocol": "cooperbench-four-arm-1.13-r5-dashscope-qwen35-cache-workspace-recovery-1",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "floating alias; backend version unavailable"
    },
    "cdb708d139b20b27": {
      "agent": "mini_swe_agent_v2",
      "catalog_canonical_slug": "deepseek/deepseek-v4-flash-20260423",
      "endpoint": "https://openrouter.ai/api/v1",
      "expected_provider": "GMICloud",
      "global_concurrency": 10,
      "infrastructure_retries": 0,
      "max_logical_llm_calls": 300,
      "max_tokens": 65536,
      "model": "deepseek/deepseek-v4-flash",
      "model_profile_id": "openrouter-deepseek-v4-flash-0423-gmicloud-fp8-thinking-high",
      "reasoning_effort": "high",
      "response_model": "deepseek/deepseek-v4-flash",
      "source_protocol": "cooperbench-three-arm-77-1.6-openrouter-gmicloud-fp8-thinking-high",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "OpenRouter dated slug and fixed GMICloud FP8 route; immutable backend weights unverified"
    },
    "d1bb88bde286528f": {
      "agent": "mini_swe_agent_v2",
      "budget_exhaustion_workspace_recovery": true,
      "empty_submission_workspace_recovery": true,
      "endpoint": "https://api.apiyi.com/v1",
      "expected_provider": null,
      "global_concurrency": 16,
      "infrastructure_retries": 3,
      "max_logical_llm_calls": 300,
      "model": "gemini-3.1-flash-lite",
      "model_profile_id": "apiyi-gemini-3.1-flash-lite-thinking-medium",
      "reasoning_effort": "medium",
      "response_model": "gemini-3.1-flash-lite",
      "source_protocol": "cooperbench-four-arm-1.24-r5-apiyi-gemini31-flash-lite-thinking-medium",
      "temperature": 0,
      "thinking": {
        "type": "enabled"
      },
      "weight_identity_status": "APIYI gemini-3.1-flash-lite stable alias; immutable upstream weights are not independently attested"
    }
  },
  "contribution_categories": [
    "authored_tests",
    "changelog",
    "docs",
    "inline_change",
    "style",
    "typing"
  ],
  "correction": {
    "certification": "hash_matched_rescore",
    "condition": "codewiki",
    "evidence": "evidence/deepseek/RESULTS.json",
    "functional_success": true,
    "historical_functional_success": false,
    "model": "deepseek",
    "patch_sha256": "5c0d8f9b2390ce3acd4a58ea1b8c6887eef3cf1a79af33320eef86fb60062376",
    "target_test_sha256": "576c1cbf0a029f7555208cb9c7ef93ca00858377cda787f83e3986acb52054ee",
    "task_id": "JIN1559-F07"
  },
  "dataset": "SYSUSELab/RepoNormBench",
  "dataset_manifest_sha256": "27973b3df998e65a7a09bedad53ef1900e799dd6d00496f6b6db2d5568bb4fc9",
  "dataset_release": "1.0.0-rc1",
  "dataset_revision": "db1c44ca2e1c67f7bdcc10ae7ff1a9176b75c29f",
  "entries": [
    {
      "condition": "raw",
      "configuration_counts": {
        "98be4ab55601bb40": 44,
        "9e7f1bbb272dd39e": 77
      },
      "functional_pass": 72,
      "functional_rate": 0.5950413223140496,
      "functional_subset": {
        "coverage": 1.0,
        "included_tasks": 72,
        "micro": 0.6523887973640856,
        "norm_applicable": 607,
        "norm_decided": 607,
        "norm_pass": 396,
        "status_counts": {
          "fail": 211,
          "pass": 396
        },
        "task_macro": 0.6667271771438438,
        "task_macro_fraction": [
          1729885,
          2594592
        ],
        "tasks_with_instances": 72
      },
      "id": "deepseek-raw-all",
      "joint_pass": 4,
      "joint_rate": 0.03305785123966942,
      "method_label": "Raw",
      "model": "deepseek",
      "model_label": "DeepSeek V4 Flash",
      "scopes": {
        "contribution": {
          "coverage": 0.9883040935672515,
          "included_tasks": 120,
          "micro": 0.32938856015779094,
          "norm_applicable": 513,
          "norm_decided": 507,
          "norm_pass": 167,
          "status_counts": {
            "fail": 340,
            "pass": 167,
            "unavailable": 6
          },
          "task_macro": 0.33819444444444446,
          "task_macro_fraction": [
            487,
            1440
          ],
          "tasks_with_instances": 121
        },
        "overall": {
          "coverage": 0.9882697947214076,
          "included_tasks": 120,
          "micro": 0.645895153313551,
          "norm_applicable": 1023,
          "norm_decided": 1011,
          "norm_pass": 653,
          "status_counts": {
            "fail": 358,
            "not_applicable": 1,
            "pass": 653,
            "unavailable": 12
          },
          "task_macro": 0.6557761451511451,
          "task_macro_fraction": [
            28357859,
            43243200
          ],
          "tasks_with_instances": 121
        },
        "prompt_omitted": {
          "coverage": 0.9895366218236173,
          "included_tasks": 120,
          "micro": 0.49697885196374625,
          "norm_applicable": 669,
          "norm_decided": 662,
          "norm_pass": 329,
          "status_counts": {
            "fail": 333,
            "not_applicable": 1,
            "pass": 329,
            "unavailable": 7
          },
          "task_macro": 0.5207488807488807,
          "task_macro_fraction": [
            31987,
            61425
          ],
          "tasks_with_instances": 121
        }
      },
      "tasks": 121
    },
    {
      "condition": "codewiki",
      "configuration_counts": {
        "410a99f7395e143f": 44,
        "9e7f1bbb272dd39e": 77
      },
      "functional_pass": 82,
      "functional_rate": 0.6776859504132231,
      "functional_subset": {
        "coverage": 1.0,
        "included_tasks": 82,
        "micro": 0.6469740634005764,
        "norm_applicable": 694,
        "norm_decided": 694,
        "norm_pass": 449,
        "status_counts": {
          "fail": 245,
          "pass": 449
        },
        "task_macro": 0.6623547184522794,
        "task_macro_fraction": [
          271837,
          410410
        ],
        "tasks_with_instances": 82
      },
      "id": "deepseek-codewiki-all",
      "joint_pass": 5,
      "joint_rate": 0.04132231404958678,
      "method_label": "CodeWiki",
      "model": "deepseek",
      "model_label": "DeepSeek V4 Flash",
      "scopes": {
        "contribution": {
          "coverage": 1.0,
          "included_tasks": 121,
          "micro": 0.33528265107212474,
          "norm_applicable": 513,
          "norm_decided": 513,
          "norm_pass": 172,
          "status_counts": {
            "fail": 341,
            "pass": 172
          },
          "task_macro": 0.3475206611570248,
          "task_macro_fraction": [
            841,
            2420
          ],
          "tasks_with_instances": 121
        },
        "overall": {
          "coverage": 1.0,
          "included_tasks": 121,
          "micro": 0.6422287390029325,
          "norm_applicable": 1023,
          "norm_decided": 1023,
          "norm_pass": 657,
          "status_counts": {
            "fail": 366,
            "not_applicable": 1,
            "pass": 657
          },
          "task_macro": 0.6565921222946016,
          "task_macro_fraction": [
            14314877,
            21801780
          ],
          "tasks_with_instances": 121
        },
        "prompt_omitted": {
          "coverage": 1.0,
          "included_tasks": 121,
          "micro": 0.4977578475336323,
          "norm_applicable": 669,
          "norm_decided": 669,
          "norm_pass": 333,
          "status_counts": {
            "fail": 336,
            "not_applicable": 1,
            "pass": 333
          },
          "task_macro": 0.5241006468279196,
          "task_macro_fraction": [
            1038757,
            1981980
          ],
          "tasks_with_instances": 121
        }
      },
      "tasks": 121
    },
    {
      "condition": "reponorm",
      "configuration_counts": {
        "5da24dad69af9ac5": 2,
        "7f1aa28d8824cf73": 44,
        "cdb708d139b20b27": 75
      },
      "functional_pass": 83,
      "functional_rate": 0.6859504132231405,
      "functional_subset": {
        "coverage": 1.0,
        "included_tasks": 83,
        "micro": 0.7471264367816092,
        "norm_applicable": 696,
        "norm_decided": 696,
        "norm_pass": 520,
        "status_counts": {
          "fail": 176,
          "pass": 520
        },
        "task_macro": 0.7621625362589218,
        "task_macro_fraction": [
          253291,
          332332
        ],
        "tasks_with_instances": 83
      },
      "id": "deepseek-reponorm-all",
      "joint_pass": 10,
      "joint_rate": 0.08264462809917356,
      "method_label": "RepoNorm (Accepted-Only)",
      "model": "deepseek",
      "model_label": "DeepSeek V4 Flash",
      "scopes": {
        "contribution": {
          "coverage": 1.0,
          "included_tasks": 121,
          "micro": 0.46393762183235865,
          "norm_applicable": 513,
          "norm_decided": 513,
          "norm_pass": 238,
          "status_counts": {
            "fail": 275,
            "pass": 238
          },
          "task_macro": 0.4918732782369146,
          "task_macro_fraction": [
            3571,
            7260
          ],
          "tasks_with_instances": 121
        },
        "overall": {
          "coverage": 1.0,
          "included_tasks": 121,
          "micro": 0.7106549364613881,
          "norm_applicable": 1023,
          "norm_decided": 1023,
          "norm_pass": 727,
          "status_counts": {
            "fail": 296,
            "not_applicable": 1,
            "pass": 727
          },
          "task_macro": 0.7264236452252981,
          "task_macro_fraction": [
            1508317,
            2076360
          ],
          "tasks_with_instances": 121
        },
        "prompt_omitted": {
          "coverage": 1.0,
          "included_tasks": 121,
          "micro": 0.5889387144992526,
          "norm_applicable": 669,
          "norm_decided": 669,
          "norm_pass": 394,
          "status_counts": {
            "fail": 275,
            "not_applicable": 1,
            "pass": 394
          },
          "task_macro": 0.6128447310265492,
          "task_macro_fraction": [
            202441,
            330330
          ],
          "tasks_with_instances": 121
        }
      },
      "tasks": 121
    },
    {
      "condition": "raw",
      "configuration_counts": {
        "106e78241d44d994": 55,
        "2620e3b66aead1c2": 40,
        "56c7cdea8073681a": 22,
        "b62eb336168a3a60": 4
      },
      "functional_pass": 54,
      "functional_rate": 0.4462809917355372,
      "functional_subset": {
        "coverage": 1.0,
        "included_tasks": 54,
        "micro": 0.6651785714285714,
        "norm_applicable": 448,
        "norm_decided": 448,
        "norm_pass": 298,
        "status_counts": {
          "fail": 150,
          "pass": 298
        },
        "task_macro": 0.6773769440436107,
        "task_macro_fraction": [
          21124,
          31185
        ],
        "tasks_with_instances": 54
      },
      "id": "qwen-raw-all",
      "joint_pass": 3,
      "joint_rate": 0.024793388429752067,
      "method_label": "Raw",
      "model": "qwen",
      "model_label": "Qwen3.5-Flash",
      "scopes": {
        "contribution": {
          "coverage": 0.9883040935672515,
          "included_tasks": 120,
          "micro": 0.3214990138067061,
          "norm_applicable": 513,
          "norm_decided": 507,
          "norm_pass": 163,
          "status_counts": {
            "fail": 344,
            "pass": 163,
            "unavailable": 6
          },
          "task_macro": 0.3215277777777778,
          "task_macro_fraction": [
            463,
            1440
          ],
          "tasks_with_instances": 121
        },
        "overall": {
          "coverage": 0.989247311827957,
          "included_tasks": 120,
          "micro": 0.633399209486166,
          "norm_applicable": 1023,
          "norm_decided": 1012,
          "norm_pass": 641,
          "status_counts": {
            "fail": 371,
            "not_applicable": 1,
            "pass": 641,
            "unavailable": 11
          },
          "task_macro": 0.6392528767528768,
          "task_macro_fraction": [
            1382167,
            2162160
          ],
          "tasks_with_instances": 121
        },
        "prompt_omitted": {
          "coverage": 0.9895366218236173,
          "included_tasks": 120,
          "micro": 0.48036253776435045,
          "norm_applicable": 669,
          "norm_decided": 662,
          "norm_pass": 318,
          "status_counts": {
            "fail": 344,
            "not_applicable": 1,
            "pass": 318,
            "unavailable": 7
          },
          "task_macro": 0.5012497456247457,
          "task_macro_fraction": [
            1970513,
            3931200
          ],
          "tasks_with_instances": 121
        }
      },
      "tasks": 121
    },
    {
      "condition": "codewiki",
      "configuration_counts": {
        "01c1c65e291a5f5a": 66,
        "2620e3b66aead1c2": 41,
        "56c7cdea8073681a": 10,
        "730849cc0bbbaf9e": 1,
        "b62eb336168a3a60": 3
      },
      "functional_pass": 54,
      "functional_rate": 0.4462809917355372,
      "functional_subset": {
        "coverage": 1.0,
        "included_tasks": 54,
        "micro": 0.6588983050847458,
        "norm_applicable": 472,
        "norm_decided": 472,
        "norm_pass": 311,
        "status_counts": {
          "fail": 161,
          "pass": 311
        },
        "task_macro": 0.6706757234535012,
        "task_macro_fraction": [
          6525487,
          9729720
        ],
        "tasks_with_instances": 54
      },
      "id": "qwen-codewiki-all",
      "joint_pass": 1,
      "joint_rate": 0.008264462809917356,
      "method_label": "CodeWiki",
      "model": "qwen",
      "model_label": "Qwen3.5-Flash",
      "scopes": {
        "contribution": {
          "coverage": 1.0,
          "included_tasks": 121,
          "micro": 0.2982456140350877,
          "norm_applicable": 513,
          "norm_decided": 513,
          "norm_pass": 153,
          "status_counts": {
            "fail": 360,
            "pass": 153
          },
          "task_macro": 0.2931129476584022,
          "task_macro_fraction": [
            532,
            1815
          ],
          "tasks_with_instances": 121
        },
        "overall": {
          "coverage": 1.0,
          "included_tasks": 121,
          "micro": 0.611328125,
          "norm_applicable": 1024,
          "norm_decided": 1024,
          "norm_pass": 626,
          "status_counts": {
            "fail": 398,
            "pass": 626
          },
          "task_macro": 0.6218421156437686,
          "task_macro_fraction": [
            2711453,
            4360356
          ],
          "tasks_with_instances": 121
        },
        "prompt_omitted": {
          "coverage": 1.0,
          "included_tasks": 121,
          "micro": 0.46417910447761196,
          "norm_applicable": 670,
          "norm_decided": 670,
          "norm_pass": 311,
          "status_counts": {
            "fail": 359,
            "pass": 311
          },
          "task_macro": 0.48532603759876486,
          "task_macro_fraction": [
            213757,
            440440
          ],
          "tasks_with_instances": 121
        }
      },
      "tasks": 121
    },
    {
      "condition": "reponorm",
      "configuration_counts": {
        "106e78241d44d994": 69,
        "2620e3b66aead1c2": 43,
        "56c7cdea8073681a": 8,
        "b62eb336168a3a60": 1
      },
      "functional_pass": 63,
      "functional_rate": 0.5206611570247934,
      "functional_subset": {
        "coverage": 1.0,
        "included_tasks": 63,
        "micro": 0.7213438735177866,
        "norm_applicable": 506,
        "norm_decided": 506,
        "norm_pass": 365,
        "status_counts": {
          "fail": 141,
          "pass": 365
        },
        "task_macro": 0.736051426527617,
        "task_macro_fraction": [
          835517,
          1135134
        ],
        "tasks_with_instances": 63
      },
      "id": "qwen-reponorm-all",
      "joint_pass": 6,
      "joint_rate": 0.049586776859504134,
      "method_label": "RepoNorm (Accepted-Only)",
      "model": "qwen",
      "model_label": "Qwen3.5-Flash",
      "scopes": {
        "contribution": {
          "coverage": 1.0,
          "included_tasks": 121,
          "micro": 0.4093567251461988,
          "norm_applicable": 513,
          "norm_decided": 513,
          "norm_pass": 210,
          "status_counts": {
            "fail": 303,
            "pass": 210
          },
          "task_macro": 0.43388429752066116,
          "task_macro_fraction": [
            105,
            242
          ],
          "tasks_with_instances": 121
        },
        "overall": {
          "coverage": 1.0,
          "included_tasks": 121,
          "micro": 0.6728515625,
          "norm_applicable": 1024,
          "norm_decided": 1024,
          "norm_pass": 689,
          "status_counts": {
            "fail": 335,
            "pass": 689
          },
          "task_macro": 0.6867111997277286,
          "task_macro_fraction": [
            4277579,
            6229080
          ],
          "tasks_with_instances": 121
        },
        "prompt_omitted": {
          "coverage": 1.0,
          "included_tasks": 121,
          "micro": 0.5328358208955224,
          "norm_applicable": 670,
          "norm_decided": 670,
          "norm_pass": 357,
          "status_counts": {
            "fail": 313,
            "pass": 357
          },
          "task_macro": 0.5581105762923945,
          "task_macro_fraction": [
            276541,
            495495
          ],
          "tasks_with_instances": 121
        }
      },
      "tasks": 121
    },
    {
      "condition": "raw",
      "configuration_counts": {
        "825f55f8bb93fff6": 77,
        "d1bb88bde286528f": 44
      },
      "functional_pass": 37,
      "functional_rate": 0.30578512396694213,
      "functional_subset": {
        "coverage": 0.9965397923875432,
        "included_tasks": 37,
        "micro": 0.65625,
        "norm_applicable": 289,
        "norm_decided": 288,
        "norm_pass": 189,
        "status_counts": {
          "error": 1,
          "fail": 99,
          "pass": 189
        },
        "task_macro": 0.670981045981046,
        "task_macro_fraction": [
          15293,
          22792
        ],
        "tasks_with_instances": 37
      },
      "id": "gemini-raw-all",
      "joint_pass": 1,
      "joint_rate": 0.008264462809917356,
      "method_label": "Raw",
      "model": "gemini",
      "model_label": "Gemini 3.1 Flash-Lite",
      "scopes": {
        "contribution": {
          "coverage": 0.9746588693957114,
          "included_tasks": 119,
          "micro": 0.264,
          "norm_applicable": 513,
          "norm_decided": 500,
          "norm_pass": 132,
          "status_counts": {
            "error": 9,
            "fail": 368,
            "pass": 132,
            "unavailable": 4
          },
          "task_macro": 0.25280112044817926,
          "task_macro_fraction": [
            361,
            1428
          ],
          "tasks_with_instances": 121
        },
        "overall": {
          "coverage": 0.97265625,
          "included_tasks": 120,
          "micro": 0.570281124497992,
          "norm_applicable": 1024,
          "norm_decided": 996,
          "norm_pass": 568,
          "status_counts": {
            "error": 22,
            "fail": 428,
            "pass": 568,
            "unavailable": 6
          },
          "task_macro": 0.5760623404373404,
          "task_macro_fraction": [
            8303593,
            14414400
          ],
          "tasks_with_instances": 121
        },
        "prompt_omitted": {
          "coverage": 0.9761194029850746,
          "included_tasks": 120,
          "micro": 0.41743119266055045,
          "norm_applicable": 670,
          "norm_decided": 654,
          "norm_pass": 273,
          "status_counts": {
            "error": 12,
            "fail": 381,
            "pass": 273,
            "unavailable": 4
          },
          "task_macro": 0.43356456043956043,
          "task_macro_fraction": [
            63127,
            145600
          ],
          "tasks_with_instances": 121
        }
      },
      "tasks": 121
    },
    {
      "condition": "codewiki",
      "configuration_counts": {
        "825f55f8bb93fff6": 77,
        "d1bb88bde286528f": 44
      },
      "functional_pass": 45,
      "functional_rate": 0.371900826446281,
      "functional_subset": {
        "coverage": 0.9945652173913043,
        "included_tasks": 45,
        "micro": 0.6584699453551912,
        "norm_applicable": 368,
        "norm_decided": 366,
        "norm_pass": 241,
        "status_counts": {
          "error": 2,
          "fail": 125,
          "pass": 241
        },
        "task_macro": 0.6738710672044005,
        "task_macro_fraction": [
          2731907,
          4054050
        ],
        "tasks_with_instances": 45
      },
      "id": "gemini-codewiki-all",
      "joint_pass": 1,
      "joint_rate": 0.008264462809917356,
      "method_label": "CodeWiki",
      "model": "gemini",
      "model_label": "Gemini 3.1 Flash-Lite",
      "scopes": {
        "contribution": {
          "coverage": 0.9727095516569201,
          "included_tasks": 120,
          "micro": 0.2685370741482966,
          "norm_applicable": 513,
          "norm_decided": 499,
          "norm_pass": 134,
          "status_counts": {
            "error": 14,
            "fail": 365,
            "pass": 134
          },
          "task_macro": 0.25625,
          "task_macro_fraction": [
            41,
            160
          ],
          "tasks_with_instances": 121
        },
        "overall": {
          "coverage": 0.986328125,
          "included_tasks": 121,
          "micro": 0.5801980198019802,
          "norm_applicable": 1024,
          "norm_decided": 1010,
          "norm_pass": 586,
          "status_counts": {
            "error": 14,
            "fail": 424,
            "pass": 586
          },
          "task_macro": 0.5835999400048987,
          "task_macro_fraction": [
            1696469,
            2906904
          ],
          "tasks_with_instances": 121
        },
        "prompt_omitted": {
          "coverage": 0.9791044776119403,
          "included_tasks": 121,
          "micro": 0.41310975609756095,
          "norm_applicable": 670,
          "norm_decided": 656,
          "norm_pass": 271,
          "status_counts": {
            "error": 14,
            "fail": 385,
            "pass": 271
          },
          "task_macro": 0.42791577109758927,
          "task_macro_fraction": [
            1696241,
            3963960
          ],
          "tasks_with_instances": 121
        }
      },
      "tasks": 121
    },
    {
      "condition": "reponorm",
      "configuration_counts": {
        "825f55f8bb93fff6": 77,
        "d1bb88bde286528f": 44
      },
      "functional_pass": 44,
      "functional_rate": 0.36363636363636365,
      "functional_subset": {
        "coverage": 0.9971509971509972,
        "included_tasks": 44,
        "micro": 0.7,
        "norm_applicable": 351,
        "norm_decided": 350,
        "norm_pass": 245,
        "status_counts": {
          "error": 1,
          "fail": 105,
          "pass": 245
        },
        "task_macro": 0.7166141938869212,
        "task_macro_fraction": [
          21851,
          30492
        ],
        "tasks_with_instances": 44
      },
      "id": "gemini-reponorm-all",
      "joint_pass": 2,
      "joint_rate": 0.01652892561983471,
      "method_label": "RepoNorm (Accepted-Only)",
      "model": "gemini",
      "model_label": "Gemini 3.1 Flash-Lite",
      "scopes": {
        "contribution": {
          "coverage": 0.9727095516569201,
          "included_tasks": 120,
          "micro": 0.3406813627254509,
          "norm_applicable": 513,
          "norm_decided": 499,
          "norm_pass": 170,
          "status_counts": {
            "error": 9,
            "fail": 329,
            "pass": 170,
            "unavailable": 5
          },
          "task_macro": 0.3327777777777778,
          "task_macro_fraction": [
            599,
            1800
          ],
          "tasks_with_instances": 121
        },
        "overall": {
          "coverage": 0.9814453125,
          "included_tasks": 120,
          "micro": 0.6199004975124378,
          "norm_applicable": 1024,
          "norm_decided": 1005,
          "norm_pass": 623,
          "status_counts": {
            "error": 11,
            "fail": 382,
            "pass": 623,
            "unavailable": 8
          },
          "task_macro": 0.6338279313279314,
          "task_macro_fraction": [
            6852187,
            10810800
          ],
          "tasks_with_instances": 121
        },
        "prompt_omitted": {
          "coverage": 0.9776119402985075,
          "included_tasks": 120,
          "micro": 0.4717557251908397,
          "norm_applicable": 670,
          "norm_decided": 655,
          "norm_pass": 309,
          "status_counts": {
            "error": 9,
            "fail": 346,
            "pass": 309,
            "unavailable": 6
          },
          "task_macro": 0.49682387057387056,
          "task_macro_fraction": [
            325519,
            655200
          ],
          "tasks_with_instances": 121
        }
      },
      "tasks": 121
    }
  ],
  "generated_by": "scripts/build_data.py",
  "generator_sha256": "0f6d948a922620540045a4b4c4015f38308f6bdd0a14cbb1c53501679aa86f28",
  "norm_profile": "coarse_auto",
  "reproduction_code_sha256": "ba9d05ba8541f3c5084f32e13e606ea634ffe0a39823c34d0225dfd5989af149",
  "schema_version": "leaderboard-1.0"
}