{
  "runId": "2026-06-05T16:44:34.030Z-qwen2.5-coder_14b",
  "date": "2026-06-05",
  "model": {
    "id": "qwen2.5-coder:14b",
    "name": "qwen2.5-coder:14b",
    "kind": "local",
    "runner": "ollama",
    "quant": ""
  },
  "env": {
    "gpu": "NVIDIA GeForce RTX 3090",
    "envClass": "gaming",
    "ollamaVersion": "ollama version is 0.19.0",
    "driver": "610.47",
    "powerLimitW": 420,
    "host": "win32 10.0.26200",
    "note": ""
  },
  "params": {
    "temperature": 0,
    "seed": 7,
    "numCtx": 8192,
    "topP": 1
  },
  "tasks": [
    {
      "id": "code-hozen-01",
      "category": "code-hozen",
      "categoryLabel": "日本語コード保守",
      "type": "agent",
      "score01": 0.3,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "script",
          "score": 1,
          "details": {
            "code": 0,
            "stdout": "LOCK OK: test.mjs は改変されていません\n",
            "stderr": ""
          }
        },
        {
          "kind": "objective",
          "method": "script",
          "score": 0,
          "details": {
            "code": 1,
            "stdout": "",
            "stderr": "node:internal/modules/run_main:107\r\n    triggerUncaughtException(\r\n    ^\r\n\r\nAssertionError [ERR_ASSERTION]: ケース2 切り捨て(0.6): expected {\"subtotal\":1995,\"tax\":159,\"total\":2154} but got {\"subtotal\":1995,\"tax\":160,\"total\":2155}\r\n+ actual - expected\r\n\r\n  {\r\n    subtotal: 1995,\r\n+   tax: 160,\r\n+   total: 2155\r\n-   tax: 159,\r\n-   total: 2154\r\n  }\r\n\r\n    at check (file:///<TMP>/jworkbench-code-hozen-01-9szT6T/fixtures/test.mjs:8:10)\r\n    at file:///<TMP>/jworkbench-code-hozen-01-9szT6T/fixtures/test.mjs:21:1\r\n    at ModuleJob.run (node:internal/modules/esm/module_job:371:25)\r\n    at async onImport.tracePromise.__proto__ (node:internal/modules/esm/loader:702:26)\r\n    at async asyncRunEntryPointWithESMLoader (node:internal/modules/run_main:101:5) {\r\n  generatedMessage: false,\r\n  code: 'ERR_ASSERTION',\r\n  actual: { subtotal: 1995, tax: 160, total: 2155 },\r\n  expected: { subtotal: 1995, tax: 159, total: 2154 },\r\n  operator: 'deepStrictEqual',\r\n  diff: 'simple'\r\n}\r\n\r\nNode.js v24.8.0\r\n"
          }
        }
      ],
      "axes": {
        "quality": 0.3
      },
      "metrics": {
        "ttftMs": null,
        "tokens": 0,
        "tokPerSec": null,
        "wallMs": 6741,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.3,
          "wallMs": 6741
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/code-hozen-01.md"
    },
    {
      "id": "code-hozen-02",
      "category": "code-hozen",
      "categoryLabel": "日本語コード保守",
      "type": "agent",
      "score01": 0.3,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "script",
          "score": 1,
          "details": {
            "code": 0,
            "stdout": "LOCK OK: validate.mjs は改変されていません\n",
            "stderr": ""
          }
        },
        {
          "kind": "objective",
          "method": "script",
          "score": 0,
          "details": {
            "code": 1,
            "stdout": "",
            "stderr": "VALIDATE FAIL: 必須キー 'description' がありません\n"
          }
        }
      ],
      "axes": {
        "quality": 0.3
      },
      "metrics": {
        "ttftMs": null,
        "tokens": 0,
        "tokPerSec": null,
        "wallMs": 1301,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.3,
          "wallMs": 1301
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/code-hozen-02.md"
    },
    {
      "id": "code-hozen-03",
      "category": "code-hozen",
      "categoryLabel": "日本語コード保守",
      "type": "agent",
      "score01": 0.2,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "script",
          "score": 1,
          "details": {
            "code": 0,
            "stdout": "LOCK OK: test.mjs は改変されていません\n",
            "stderr": ""
          }
        },
        {
          "kind": "objective",
          "method": "script",
          "score": 0,
          "details": {
            "code": 1,
            "stdout": "",
            "stderr": "<TMP>\\jworkbench-code-hozen-03-ZnAE3q\\fixtures\\broken.mjs:22\r\n  if (inRangeInclusive(qty, 1, 10) {\r\n                                   ^\r\n\r\nSyntaxError: Unexpected token '{'\r\n    at checkSyntax (node:internal/main/check_syntax:72:5)\r\n\r\nNode.js v24.8.0\r\n"
          }
        },
        {
          "kind": "objective",
          "method": "script",
          "score": 0,
          "details": {
            "code": 1,
            "stdout": "",
            "stderr": "file:///<TMP>/jworkbench-code-hozen-03-ZnAE3q/fixtures/broken.mjs:22\r\n  if (inRangeInclusive(qty, 1, 10) {\r\n                                   ^\r\n\r\nSyntaxError: Unexpected token '{'\r\n    at compileSourceTextModule (node:internal/modules/esm/utils:357:16)\r\n    at ModuleLoader.moduleStrategy (node:internal/modules/esm/translators:109:18)\r\n    at #translate (node:internal/modules/esm/loader:575:20)\r\n    at afterLoad (node:internal/modules/esm/loader:625:29)\r\n    at ModuleLoader.loadAndTranslate (node:internal/modules/esm/loader:630:12)\r\n    at #createModuleJob (node:internal/modules/esm/loader:654:36)\r\n    at #getJobFromResolveResult (node:internal/modules/esm/loader:354:34)\r\n    at ModuleLoader.getModuleJobForImport (node:internal/modules/esm/loader:319:41)\r\n\r\nNode.js v24.8.0\r\n"
          }
        }
      ],
      "axes": {
        "quality": 0.2
      },
      "metrics": {
        "ttftMs": null,
        "tokens": 0,
        "tokPerSec": null,
        "wallMs": 1219,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.2,
          "wallMs": 1219
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/code-hozen-03.md"
    },
    {
      "id": "code-hozen-04",
      "category": "code-hozen",
      "categoryLabel": "日本語コード保守",
      "type": "agent",
      "score01": 0.3,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "script",
          "score": 1,
          "details": {
            "code": 0,
            "stdout": "LOCK OK: test.mjs は改変されていません\n",
            "stderr": ""
          }
        },
        {
          "kind": "objective",
          "method": "script",
          "score": 0,
          "details": {
            "code": 1,
            "stdout": "",
            "stderr": "file:///<TMP>/jworkbench-code-hozen-04-abMs52/fixtures/summarize.mjs:15\r\n  throw new Error('TODO: daysInMonth は未実装です');\r\n        ^\r\n\r\nError: TODO: daysInMonth は未実装です\r\n    at daysInMonth (file:///<TMP>/jworkbench-code-hozen-04-abMs52/fixtures/summarize.mjs:15:9)\r\n    at file:///<TMP>/jworkbench-code-hozen-04-abMs52/fixtures/test.mjs:7:14\r\n    at ModuleJob.run (node:internal/modules/esm/module_job:371:25)\r\n    at async onImport.tracePromise.__proto__ (node:internal/modules/esm/loader:702:26)\r\n    at async asyncRunEntryPointWithESMLoader (node:internal/modules/run_main:101:5)\r\n\r\nNode.js v24.8.0\r\n"
          }
        }
      ],
      "axes": {
        "quality": 0.3
      },
      "metrics": {
        "ttftMs": null,
        "tokens": 0,
        "tokPerSec": null,
        "wallMs": 1459,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.3,
          "wallMs": 1459
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/code-hozen-04.md"
    },
    {
      "id": "gijiroku-01",
      "category": "gijiroku",
      "categoryLabel": "議事録ToDo抽出",
      "type": "qa",
      "score01": 1,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "佐倉",
              "並木",
              "桜井"
            ],
            "hits": [
              "佐倉",
              "桜井"
            ]
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "2026-04-10",
              "null",
              "2026-04-20"
            ],
            "hits": [
              "2026-04-10",
              "2026-04-20"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": 404,
        "tokens": 172,
        "tokPerSec": 71.01568951279934,
        "wallMs": 2826,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 2826
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/gijiroku-01.md"
    },
    {
      "id": "gijiroku-02",
      "category": "gijiroku",
      "categoryLabel": "議事録ToDo抽出",
      "type": "qa",
      "score01": 0.6749999999999999,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0.5,
          "details": {
            "values": [
              "2026-05-16",
              "2026-05-13",
              "2026-05-31"
            ],
            "hits": [
              "2026-05-31"
            ]
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "宮田",
              "岩本",
              "柏木"
            ],
            "hits": [
              "宮田",
              "柏木"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.7,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.6749999999999999
      },
      "metrics": {
        "ttftMs": 4694,
        "tokens": 179,
        "tokPerSec": 71.11640842272547,
        "wallMs": 7211,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.6749999999999999,
          "wallMs": 7211
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/gijiroku-02.md"
    },
    {
      "id": "gijiroku-03",
      "category": "gijiroku",
      "categoryLabel": "議事録ToDo抽出",
      "type": "qa",
      "score01": 0.6000000000000001,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "values": [],
            "hits": []
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "テンプレを入れるツールのアカウント管理者権限がないと反映できない"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.8,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.6000000000000001
      },
      "metrics": {
        "ttftMs": 4966,
        "tokens": 107,
        "tokPerSec": 71.42857142857143,
        "wallMs": 6464,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.6000000000000001,
          "wallMs": 6464
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/gijiroku-03.md"
    },
    {
      "id": "gijiroku-04",
      "category": "gijiroku",
      "categoryLabel": "議事録ToDo抽出",
      "type": "qa",
      "score01": 1,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "2026-03-04",
              "2026-03-05",
              "2026-02-20"
            ],
            "hits": [
              "2026-02-20",
              "2026-03-05"
            ]
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "神保",
              "千葉",
              "久米"
            ],
            "hits": [
              "久米",
              "千葉"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": 4723,
        "tokens": 222,
        "tokPerSec": 71.0172744721689,
        "wallMs": 7849,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 7849
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/gijiroku-04.md"
    },
    {
      "id": "gijiroku-05",
      "category": "gijiroku",
      "categoryLabel": "議事録ToDo抽出",
      "type": "qa",
      "score01": 1,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "2026-06-12",
              "2026-06-20"
            ],
            "hits": [
              "2026-06-12",
              "2026-06-20"
            ]
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "会場予約",
              "出欠フォーム作成と集計"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": 4882,
        "tokens": 121,
        "tokPerSec": 70.926143024619,
        "wallMs": 6588,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 6588
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/gijiroku-05.md"
    },
    {
      "id": "hyou-csv-01",
      "category": "hyou-csv",
      "categoryLabel": "表/CSV処理",
      "type": "qa",
      "score01": 0.04000000000000001,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "numeric",
          "score": 0,
          "details": {
            "actual": 87400
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "values": [
              "15300"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.2,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.04000000000000001
      },
      "metrics": {
        "ttftMs": 4927,
        "tokens": 98,
        "tokPerSec": 71.42857142857142,
        "wallMs": 6299,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.04000000000000001,
          "wallMs": 6299
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/hyou-csv-01.md"
    },
    {
      "id": "hyou-csv-02",
      "category": "hyou-csv",
      "categoryLabel": "表/CSV処理",
      "type": "qa",
      "score01": 0.82,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "numeric",
          "score": 1,
          "details": {
            "actual": 16800
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "true"
            ]
          }
        },
        {
          "kind": "objective",
          "method": "numeric",
          "score": 0,
          "details": {
            "actual": 37900
          }
        },
        {
          "kind": "rubric",
          "score": 0.6,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.82
      },
      "metrics": {
        "ttftMs": 4981,
        "tokens": 41,
        "tokPerSec": 71.80385288966725,
        "wallMs": 5552,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.82,
          "wallMs": 5552
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/hyou-csv-02.md"
    },
    {
      "id": "hyou-csv-03",
      "category": "hyou-csv",
      "categoryLabel": "表/CSV処理",
      "type": "qa",
      "score01": 0.64,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0.6666666666666666,
          "details": {
            "values": [
              "340",
              "200",
              "1",
              "4000",
              "88",
              "0"
            ],
            "hits": [
              4000,
              0
            ]
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "2026-04-03",
              "2026-04-03",
              "2026-04-10",
              "2026-04-12",
              "2026-04-15",
              "2026-04-03"
            ],
            "hits": [
              "2026-04-12",
              "2026-04-15",
              "2026-04-03"
            ]
          }
        },
        {
          "kind": "objective",
          "method": "numeric",
          "score": 0,
          "details": {
            "actual": 579840
          }
        },
        {
          "kind": "rubric",
          "score": 0.7,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.64
      },
      "metrics": {
        "ttftMs": 4930,
        "tokens": 353,
        "tokPerSec": 71.12633487809792,
        "wallMs": 9893,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.64,
          "wallMs": 9893
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/hyou-csv-03.md"
    },
    {
      "id": "hyou-csv-04",
      "category": "hyou-csv",
      "categoryLabel": "表/CSV処理",
      "type": "qa",
      "score01": 0.54,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "R-01",
              "R-02",
              "R-03",
              "R-04",
              "R-05",
              "R-06",
              "R-07"
            ],
            "hits": [
              "R-03",
              "R-07"
            ]
          }
        },
        {
          "kind": "objective",
          "method": "numeric",
          "score": 0,
          "details": {
            "actual": 860000
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "false"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.2,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.54
      },
      "metrics": {
        "ttftMs": 4655,
        "tokens": 284,
        "tokPerSec": 71.51850919163938,
        "wallMs": 8626,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.54,
          "wallMs": 8626
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/hyou-csv-04.md"
    },
    {
      "id": "keiyaku-01",
      "category": "keiyaku",
      "categoryLabel": "契約・見積照合",
      "type": "qa",
      "score01": 0.35,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "values": [
              "3"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.7,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.35
      },
      "metrics": {
        "ttftMs": 4708,
        "tokens": 153,
        "tokPerSec": 71.06363214119833,
        "wallMs": 6861,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.35,
          "wallMs": 6861
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/keiyaku-01.md"
    },
    {
      "id": "keiyaku-02",
      "category": "keiyaku",
      "categoryLabel": "契約・見積照合",
      "type": "qa",
      "score01": 0.15,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "values": [
              "640200"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.3,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.15
      },
      "metrics": {
        "ttftMs": 4894,
        "tokens": 108,
        "tokPerSec": 71.6655607166556,
        "wallMs": 6401,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.15,
          "wallMs": 6401
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/keiyaku-02.md"
    },
    {
      "id": "keiyaku-03",
      "category": "keiyaku",
      "categoryLabel": "契約・見積照合",
      "type": "qa",
      "score01": 0.15,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "values": [
              "1"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.3,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "reliability": 0.15
      },
      "metrics": {
        "ttftMs": 4994,
        "tokens": 353,
        "tokPerSec": 70.3187250996016,
        "wallMs": 10014,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.15,
          "wallMs": 10014
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/keiyaku-03.md"
    },
    {
      "id": "keiyaku-04",
      "category": "keiyaku",
      "categoryLabel": "契約・見積照合",
      "type": "qa",
      "score01": 1,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "金10,000,000円"
            ],
            "hits": [
              "10,000,000"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": 5087,
        "tokens": 149,
        "tokPerSec": 70.38261691072272,
        "wallMs": 7204,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 7204
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/keiyaku-04.md"
    },
    {
      "id": "keiyaku-05",
      "category": "keiyaku",
      "categoryLabel": "契約・見積照合",
      "type": "qa",
      "score01": 0.7,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "includes",
          "score": 1,
          "details": {}
        },
        {
          "kind": "rubric",
          "score": 0.5,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.7
      },
      "metrics": {
        "ttftMs": 4747,
        "tokens": 571,
        "tokPerSec": 70.0699472327893,
        "wallMs": 12896,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.7,
          "wallMs": 12896
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/keiyaku-05.md"
    },
    {
      "id": "longctx-01",
      "category": "longctx",
      "categoryLabel": "長文コンテキスト耐性",
      "type": "qa",
      "score01": 1,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "includes",
          "score": 1,
          "details": {}
        },
        {
          "kind": "objective",
          "method": "numeric",
          "score": 1,
          "details": {
            "actual": 4
          }
        },
        {
          "kind": "objective",
          "method": "regex",
          "score": 1,
          "details": {}
        }
      ],
      "axes": {
        "reliability": 1
      },
      "metrics": {
        "ttftMs": 5367,
        "tokens": 57,
        "tokPerSec": 68.26347305389221,
        "wallMs": 6202,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 6202
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/longctx-01.md"
    },
    {
      "id": "longctx-02",
      "category": "longctx",
      "categoryLabel": "長文コンテキスト耐性",
      "type": "qa",
      "score01": 1,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "v4.2.0"
            ]
          }
        },
        {
          "kind": "objective",
          "method": "includes",
          "score": 1,
          "details": {}
        },
        {
          "kind": "objective",
          "method": "numeric",
          "score": 1,
          "details": {
            "actual": 11000
          }
        }
      ],
      "axes": {
        "reliability": 1
      },
      "metrics": {
        "ttftMs": 1128,
        "tokens": 56,
        "tokPerSec": 67.96116504854369,
        "wallMs": 1952,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 1952
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/longctx-02.md"
    },
    {
      "id": "mail-01",
      "category": "mail",
      "categoryLabel": "メール返信・敬語",
      "type": "qa",
      "score01": 0.7,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "includes",
          "score": 0,
          "details": {}
        },
        {
          "kind": "objective",
          "method": "includes",
          "score": 1,
          "details": {}
        },
        {
          "kind": "rubric",
          "score": 0.9,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "reliability": 0.7
      },
      "metrics": {
        "ttftMs": 430,
        "tokens": 355,
        "tokPerSec": 70.43650793650794,
        "wallMs": 5470,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.7,
          "wallMs": 5470
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/mail-01.md"
    },
    {
      "id": "mail-02",
      "category": "mail",
      "categoryLabel": "メール返信・敬語",
      "type": "qa",
      "score01": 0.95,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "includes",
          "score": 1,
          "details": {}
        },
        {
          "kind": "objective",
          "method": "includes",
          "score": 1,
          "details": {}
        },
        {
          "kind": "rubric",
          "score": 0.9,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "reliability": 0.95
      },
      "metrics": {
        "ttftMs": 4746,
        "tokens": 313,
        "tokPerSec": 70.52726453357369,
        "wallMs": 9184,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.95,
          "wallMs": 9184
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/mail-02.md"
    },
    {
      "id": "mail-03",
      "category": "mail",
      "categoryLabel": "メール返信・敬語",
      "type": "qa",
      "score01": 0.95,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "includes",
          "score": 1,
          "details": {}
        },
        {
          "kind": "objective",
          "method": "includes",
          "score": 1,
          "details": {}
        },
        {
          "kind": "rubric",
          "score": 0.9,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "reliability": 0.95
      },
      "metrics": {
        "ttftMs": 4985,
        "tokens": 254,
        "tokPerSec": 70.3991130820399,
        "wallMs": 8593,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.95,
          "wallMs": 8593
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/mail-03.md"
    },
    {
      "id": "mail-04",
      "category": "mail",
      "categoryLabel": "メール返信・敬語",
      "type": "qa",
      "score01": 1,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "includes",
          "score": 1,
          "details": {}
        },
        {
          "kind": "objective",
          "method": "includes",
          "score": 1,
          "details": {}
        },
        {
          "kind": "rubric",
          "score": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "reliability": 1
      },
      "metrics": {
        "ttftMs": 4936,
        "tokens": 245,
        "tokPerSec": 70.64590542099192,
        "wallMs": 8404,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 8404
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/mail-04.md"
    },
    {
      "id": "shisya-rules-01",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 0.7,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "第5条"
            ],
            "hits": [
              "第5条"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.5,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.7
      },
      "metrics": {
        "ttftMs": 4975,
        "tokens": 274,
        "tokPerSec": 70.38273824813768,
        "wallMs": 8868,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.7,
          "wallMs": 8868
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/shisya-rules-01.md"
    },
    {
      "id": "shisya-rules-02",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 0.5,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0.5,
          "details": {
            "values": [
              "細則1",
              "細則3"
            ],
            "hits": [
              "細則3"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.5,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.5
      },
      "metrics": {
        "ttftMs": 5062,
        "tokens": 365,
        "tokPerSec": 70.46332046332047,
        "wallMs": 10242,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.5,
          "wallMs": 10242
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/shisya-rules-02.md"
    },
    {
      "id": "shisya-rules-03",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 0.7250000000000001,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "第24条"
            ],
            "hits": [
              "第24条"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.5,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.7250000000000001
      },
      "metrics": {
        "ttftMs": 5019,
        "tokens": 235,
        "tokPerSec": 70.97553609181516,
        "wallMs": 8330,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.7250000000000001,
          "wallMs": 8330
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/shisya-rules-03.md"
    },
    {
      "id": "shisya-rules-04",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 0.88,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "不可",
              "可",
              "不可",
              "不可"
            ],
            "hits": [
              "不可"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.8,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "reliability": 0.88
      },
      "metrics": {
        "ttftMs": 5051,
        "tokens": 335,
        "tokPerSec": 70.36336904011762,
        "wallMs": 9812,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.88,
          "wallMs": 9812
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/shisya-rules-04.md"
    },
    {
      "id": "shisya-rules-05",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 0.8200000000000001,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "14日",
              "10日",
              "不可",
              "可"
            ],
            "hits": [
              "14"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.7,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.8200000000000001
      },
      "metrics": {
        "ttftMs": 4983,
        "tokens": 454,
        "tokPerSec": 69.89992301770593,
        "wallMs": 11478,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.8200000000000001,
          "wallMs": 11478
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/shisya-rules-05.md"
    },
    {
      "id": "shisya-rules-06",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 0.8900000000000001,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "在宅勤務規程 第3条/第5条",
              "CSIRT運用 連絡細則 細則2",
              "情報セキュリティ規程 第10条第3項/CSIRT運用 連絡細則 細則4"
            ],
            "hits": [
              "第3条",
              "細則2",
              "第10条"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.8,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.8900000000000001
      },
      "metrics": {
        "ttftMs": 5072,
        "tokens": 465,
        "tokPerSec": 69.51711765585289,
        "wallMs": 11761,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.8900000000000001,
          "wallMs": 11761
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5-coder_14b/shisya-rules-06.md"
    }
  ],
  "rollup": {
    "byCategory": {
      "code-hozen": {
        "scoreMean": 0.275,
        "passRate": 0
      },
      "gijiroku": {
        "scoreMean": 0.8550000000000001,
        "passRate": 0.6
      },
      "hyou-csv": {
        "scoreMean": 0.51,
        "passRate": 0.25
      },
      "keiyaku": {
        "scoreMean": 0.4699999999999999,
        "passRate": 0.4
      },
      "longctx": {
        "scoreMean": 1,
        "passRate": 1
      },
      "mail": {
        "scoreMean": 0.8999999999999999,
        "passRate": 1
      },
      "shisya-rules": {
        "scoreMean": 0.7525000000000001,
        "passRate": 0.8333333333333334
      }
    },
    "byAxis": {
      "quality": 66,
      "reliability": 89,
      "speed": 100,
      "economy": 95,
      "localValue": 95,
      "setup": 70
    },
    "total100": 83,
    "capEvents": [],
    "tasksCompleted": 30,
    "tasksAttempted": 30
  },
  "rescoredAt": "2026-06-05",
  "judge": "claude (rubric, workflow-rejudge) + objective(substring)"
}