{
  "runId": "2026-06-05T16:57:30.182Z-qwen2.5_7b",
  "date": "2026-06-05",
  "model": {
    "id": "qwen2.5:7b",
    "name": "qwen2.5:7b",
    "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": 1,
      "passed": true,
      "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": 1,
          "details": {
            "code": 0,
            "stdout": "OK: 全ケース合格\n",
            "stderr": ""
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": null,
        "tokens": 0,
        "tokPerSec": null,
        "wallMs": 15772,
        "toolCalls": 6,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 15772
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/code-hozen-01.md"
    },
    {
      "id": "code-hozen-02",
      "category": "code-hozen",
      "categoryLabel": "日本語コード保守",
      "type": "agent",
      "score01": 1,
      "passed": true,
      "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": 1,
          "details": {
            "code": 0,
            "stdout": "VALIDATE OK: frontmatter は必須キーと形式を満たしています\n",
            "stderr": ""
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": null,
        "tokens": 0,
        "tokPerSec": null,
        "wallMs": 15677,
        "toolCalls": 4,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 15677
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/code-hozen-02.md"
    },
    {
      "id": "code-hozen-03",
      "category": "code-hozen",
      "categoryLabel": "日本語コード保守",
      "type": "agent",
      "score01": 1,
      "passed": true,
      "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": 1,
          "details": {
            "code": 0,
            "stdout": "",
            "stderr": ""
          }
        },
        {
          "kind": "objective",
          "method": "script",
          "score": 1,
          "details": {
            "code": 0,
            "stdout": "OK: 全ケース合格\n",
            "stderr": ""
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": null,
        "tokens": 0,
        "tokPerSec": null,
        "wallMs": 11633,
        "toolCalls": 6,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 11633
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/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-Oq1p4W/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-Oq1p4W/fixtures/summarize.mjs:15:9)\r\n    at file:///<TMP>/jworkbench-code-hozen-04-Oq1p4W/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": 194122,
        "toolCalls": 5,
        "toolCallInvalid": 3,
        "capHit": false,
        "errored": true,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.3,
          "wallMs": 194122
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/code-hozen-04.md"
    },
    {
      "id": "gijiroku-01",
      "category": "gijiroku",
      "categoryLabel": "議事録ToDo抽出",
      "type": "qa",
      "score01": 0.65,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0.5,
          "details": {
            "values": [
              "佐倉",
              "並木"
            ],
            "hits": [
              "佐倉"
            ]
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0.5,
          "details": {
            "values": [
              "2026-04-10",
              "null"
            ],
            "hits": [
              "2026-04-10"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.8,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.65
      },
      "metrics": {
        "ttftMs": 858,
        "tokens": 131,
        "tokPerSec": 132.05645161290323,
        "wallMs": 1850,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.65,
          "wallMs": 1850
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/gijiroku-01.md"
    },
    {
      "id": "gijiroku-02",
      "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": [
              "2026-05-18",
              "2026-05-17",
              "2026-05-31"
            ],
            "hits": [
              "2026-05-18",
              "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": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": 247,
        "tokens": 201,
        "tokPerSec": 133.82157123834887,
        "wallMs": 1749,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 1749
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/gijiroku-02.md"
    },
    {
      "id": "gijiroku-03",
      "category": "gijiroku",
      "categoryLabel": "議事録ToDo抽出",
      "type": "qa",
      "score01": 0.85,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "江口（CS）",
              "藤堂（CS）"
            ],
            "hits": [
              "江口"
            ]
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "権限がないとテンプレ刷新後の反映ができない可能性がある"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.7,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.85
      },
      "metrics": {
        "ttftMs": 256,
        "tokens": 162,
        "tokPerSec": 133.44316309719935,
        "wallMs": 1470,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.85,
          "wallMs": 1470
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/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": [
              "2026-03-05",
              "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": 280,
        "tokens": 221,
        "tokPerSec": 131.2351543942993,
        "wallMs": 1964,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 1964
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/gijiroku-04.md"
    },
    {
      "id": "gijiroku-05",
      "category": "gijiroku",
      "categoryLabel": "議事録ToDo抽出",
      "type": "qa",
      "score01": 0.9,
      "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": 0.8,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.9
      },
      "metrics": {
        "ttftMs": 289,
        "tokens": 134,
        "tokPerSec": 133.20079522862824,
        "wallMs": 1295,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.9,
          "wallMs": 1295
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/gijiroku-05.md"
    },
    {
      "id": "hyou-csv-01",
      "category": "hyou-csv",
      "categoryLabel": "表/CSV処理",
      "type": "qa",
      "score01": 0.06,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "numeric",
          "score": 0,
          "details": {
            "actual": 67880
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "values": [
              "23180"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.3,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.06
      },
      "metrics": {
        "ttftMs": 285,
        "tokens": 87,
        "tokPerSec": 135.09316770186334,
        "wallMs": 929,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.06,
          "wallMs": 929
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/hyou-csv-01.md"
    },
    {
      "id": "hyou-csv-02",
      "category": "hyou-csv",
      "categoryLabel": "表/CSV処理",
      "type": "qa",
      "score01": 0.04000000000000001,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "numeric",
          "score": 0,
          "details": {
            "actual": 13200
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "values": [
              "false"
            ]
          }
        },
        {
          "kind": "objective",
          "method": "numeric",
          "score": 0,
          "details": {
            "actual": 18000
          }
        },
        {
          "kind": "rubric",
          "score": 0.2,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.04000000000000001
      },
      "metrics": {
        "ttftMs": 287,
        "tokens": 37,
        "tokPerSec": 133.0935251798561,
        "wallMs": 565,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.04000000000000001,
          "wallMs": 565
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/hyou-csv-02.md"
    },
    {
      "id": "hyou-csv-03",
      "category": "hyou-csv",
      "categoryLabel": "表/CSV処理",
      "type": "qa",
      "score01": 0.76,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "340",
              "200",
              "1250",
              "4000",
              "88",
              "0"
            ],
            "hits": [
              4000,
              1250,
              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": 579880
          }
        },
        {
          "kind": "rubric",
          "score": 0.8,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.76
      },
      "metrics": {
        "ttftMs": 271,
        "tokens": 352,
        "tokPerSec": 134.65952563121652,
        "wallMs": 2885,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.76,
          "wallMs": 2885
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/hyou-csv-03.md"
    },
    {
      "id": "hyou-csv-04",
      "category": "hyou-csv",
      "categoryLabel": "表/CSV処理",
      "type": "qa",
      "score01": 0.38000000000000006,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0.5,
          "details": {
            "values": [
              "R-03"
            ],
            "hits": [
              "R-03"
            ]
          }
        },
        {
          "kind": "objective",
          "method": "numeric",
          "score": 0,
          "details": {
            "actual": 852000
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "false"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.4,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.38000000000000006
      },
      "metrics": {
        "ttftMs": 253,
        "tokens": 76,
        "tokPerSec": 133.0998248686515,
        "wallMs": 824,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.38000000000000006,
          "wallMs": 824
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/hyou-csv-04.md"
    },
    {
      "id": "keiyaku-01",
      "category": "keiyaku",
      "categoryLabel": "契約・見積照合",
      "type": "qa",
      "score01": 0.4,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "values": [
              "3"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.8,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.4
      },
      "metrics": {
        "ttftMs": 275,
        "tokens": 122,
        "tokPerSec": 133.91877058177826,
        "wallMs": 1186,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.4,
          "wallMs": 1186
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/keiyaku-01.md"
    },
    {
      "id": "keiyaku-02",
      "category": "keiyaku",
      "categoryLabel": "契約・見積照合",
      "type": "qa",
      "score01": 0.1,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "error": "answer is not valid JSON",
            "invalidJson": true
          }
        },
        {
          "kind": "rubric",
          "score": 0.2,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.1
      },
      "metrics": {
        "ttftMs": 255,
        "tokens": 236,
        "tokPerSec": 132.58426966292134,
        "wallMs": 2035,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.1,
          "wallMs": 2035
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/keiyaku-02.md"
    },
    {
      "id": "keiyaku-03",
      "category": "keiyaku",
      "categoryLabel": "契約・見積照合",
      "type": "qa",
      "score01": 0.75,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "0"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.5,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "reliability": 0.75
      },
      "metrics": {
        "ttftMs": 298,
        "tokens": 327,
        "tokPerSec": 129.1979454760964,
        "wallMs": 2829,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.75,
          "wallMs": 2829
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/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": 344,
        "tokens": 189,
        "tokPerSec": 130.61506565307533,
        "wallMs": 1791,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 1791
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/keiyaku-04.md"
    },
    {
      "id": "keiyaku-05",
      "category": "keiyaku",
      "categoryLabel": "契約・見積照合",
      "type": "qa",
      "score01": 0.5800000000000001,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "includes",
          "score": 1,
          "details": {}
        },
        {
          "kind": "rubric",
          "score": 0.3,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.5800000000000001
      },
      "metrics": {
        "ttftMs": 290,
        "tokens": 501,
        "tokPerSec": 131.1175085056268,
        "wallMs": 4111,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.5800000000000001,
          "wallMs": 4111
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/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": 622,
        "tokens": 53,
        "tokPerSec": 126.49164677804296,
        "wallMs": 1041,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 1041
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/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": 627,
        "tokens": 52,
        "tokPerSec": 126.82926829268294,
        "wallMs": 1037,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 1037
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/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": 268,
        "tokens": 319,
        "tokPerSec": 130.31045751633988,
        "wallMs": 2716,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.7,
          "wallMs": 2716
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/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": 284,
        "tokens": 257,
        "tokPerSec": 131.25638406537283,
        "wallMs": 2242,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.95,
          "wallMs": 2242
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/mail-02.md"
    },
    {
      "id": "mail-03",
      "category": "mail",
      "categoryLabel": "メール返信・敬語",
      "type": "qa",
      "score01": 0.75,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "includes",
          "score": 0,
          "details": {}
        },
        {
          "kind": "objective",
          "method": "includes",
          "score": 1,
          "details": {}
        },
        {
          "kind": "rubric",
          "score": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "reliability": 0.75
      },
      "metrics": {
        "ttftMs": 290,
        "tokens": 297,
        "tokPerSec": 129.29908576404006,
        "wallMs": 2587,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.75,
          "wallMs": 2587
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/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": 280,
        "tokens": 251,
        "tokPerSec": 130.52522100884036,
        "wallMs": 2203,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 2203
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/mail-04.md"
    },
    {
      "id": "shisya-rules-01",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 0.5800000000000001,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "第5条",
              "第7条"
            ],
            "hits": [
              "第5条"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.3,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.5800000000000001
      },
      "metrics": {
        "ttftMs": 278,
        "tokens": 136,
        "tokPerSec": 131.27413127413126,
        "wallMs": 1314,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.5800000000000001,
          "wallMs": 1314
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/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": [
              "第4条第2項(1)",
              "第9条"
            ],
            "hits": [
              "第9条"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.5,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.5
      },
      "metrics": {
        "ttftMs": 338,
        "tokens": 217,
        "tokPerSec": 128.55450236966826,
        "wallMs": 2026,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.5,
          "wallMs": 2026
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/shisya-rules-02.md"
    },
    {
      "id": "shisya-rules-03",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 0.165,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "values": [
              "第22条第3項"
            ],
            "hits": []
          }
        },
        {
          "kind": "rubric",
          "score": 0.3,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.165
      },
      "metrics": {
        "ttftMs": 332,
        "tokens": 119,
        "tokPerSec": 130.91309130913092,
        "wallMs": 1241,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.165,
          "wallMs": 1241
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/shisya-rules-03.md"
    },
    {
      "id": "shisya-rules-04",
      "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": [
              "不可",
              "可",
              "不可",
              "情報不足"
            ],
            "hits": [
              "不可"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.5,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "reliability": 0.7
      },
      "metrics": {
        "ttftMs": 314,
        "tokens": 576,
        "tokPerSec": 128.7150837988827,
        "wallMs": 4789,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.7,
          "wallMs": 4789
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/shisya-rules-04.md"
    },
    {
      "id": "shisya-rules-05",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 0.06,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "values": [
              "10日",
              "子が1人の場合: 5日 / 子が2人以上の場合は: 10日のうち3日",
              "情報不足",
              "情報不足"
            ],
            "hits": []
          }
        },
        {
          "kind": "rubric",
          "score": 0.1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.06
      },
      "metrics": {
        "ttftMs": 329,
        "tokens": 532,
        "tokPerSec": 129.56648806624452,
        "wallMs": 4435,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.06,
          "wallMs": 4435
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/shisya-rules-05.md"
    },
    {
      "id": "shisya-rules-06",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 0.315,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0.3333333333333333,
          "details": {
            "values": [
              "第3条/第5条"
            ],
            "hits": [
              "第3条"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.3,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.315
      },
      "metrics": {
        "ttftMs": 360,
        "tokens": 209,
        "tokPerSec": 129.1718170580964,
        "wallMs": 1978,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.315,
          "wallMs": 1978
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/qwen2.5_7b/shisya-rules-06.md"
    }
  ],
  "rollup": {
    "byCategory": {
      "code-hozen": {
        "scoreMean": 0.825,
        "passRate": 0.75
      },
      "gijiroku": {
        "scoreMean": 0.8800000000000001,
        "passRate": 0.8
      },
      "hyou-csv": {
        "scoreMean": 0.31,
        "passRate": 0.25
      },
      "keiyaku": {
        "scoreMean": 0.5660000000000001,
        "passRate": 0.4
      },
      "longctx": {
        "scoreMean": 1,
        "passRate": 1
      },
      "mail": {
        "scoreMean": 0.85,
        "passRate": 1
      },
      "shisya-rules": {
        "scoreMean": 0.38666666666666666,
        "passRate": 0.16666666666666666
      }
    },
    "byAxis": {
      "quality": 65,
      "reliability": 88,
      "speed": 99,
      "economy": 95,
      "localValue": 95,
      "setup": 68
    },
    "total100": 82,
    "capEvents": [],
    "tasksCompleted": 29,
    "tasksAttempted": 30
  },
  "rescoredAt": "2026-06-05",
  "judge": "claude (rubric, workflow-rejudge) + objective(substring)"
}