{
  "runId": "2026-06-05T16:27:47.624Z-gpt-oss_20b",
  "date": "2026-06-05",
  "model": {
    "id": "gpt-oss:20b",
    "name": "gpt-oss:20b",
    "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": 11519,
        "toolCalls": 3,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 11519
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/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": 5345,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.3,
          "wallMs": 5345
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/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": 7226,
        "toolCalls": 4,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 7226
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/code-hozen-03.md"
    },
    {
      "id": "code-hozen-04",
      "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": 9902,
        "toolCalls": 1,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 9902
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/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": 11412,
        "tokens": 1683,
        "tokPerSec": 1125,
        "wallMs": 12908,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 12908
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/gijiroku-01.md"
    },
    {
      "id": "gijiroku-02",
      "category": "gijiroku",
      "categoryLabel": "議事録ToDo抽出",
      "type": "qa",
      "score01": 0.725,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0.5,
          "details": {
            "values": [
              "2026-05-16",
              "2026-05-15",
              "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": 0.8,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.725
      },
      "metrics": {
        "ttftMs": 14135,
        "tokens": 1169,
        "tokPerSec": 808.4370677731674,
        "wallMs": 15581,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.725,
          "wallMs": 15581
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/gijiroku-02.md"
    },
    {
      "id": "gijiroku-03",
      "category": "gijiroku",
      "categoryLabel": "議事録ToDo抽出",
      "type": "qa",
      "score01": 0.55,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "values": [],
            "hits": []
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "Admin account not confirmed, may block template refresh",
              "Next week's inquiries may delay inventory, impacting March 31 deadline"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.7,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.55
      },
      "metrics": {
        "ttftMs": 14352,
        "tokens": 1085,
        "tokPerSec": 1787.479406919275,
        "wallMs": 14959,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.55,
          "wallMs": 14959
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/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": [
              "2026-03-03",
              "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": 17556,
        "tokens": 1674,
        "tokPerSec": 914.7540983606557,
        "wallMs": 19386,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 19386
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/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": 11784,
        "tokens": 747,
        "tokPerSec": 917.6904176904178,
        "wallMs": 12598,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 12598
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/gijiroku-05.md"
    },
    {
      "id": "hyou-csv-01",
      "category": "hyou-csv",
      "categoryLabel": "表/CSV処理",
      "type": "qa",
      "score01": 1,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "numeric",
          "score": 1,
          "details": {
            "actual": 106580
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "31980"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": 16215,
        "tokens": 1315,
        "tokPerSec": 2815.845824411135,
        "wallMs": 16682,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 16682
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/hyou-csv-01.md"
    },
    {
      "id": "hyou-csv-02",
      "category": "hyou-csv",
      "categoryLabel": "表/CSV処理",
      "type": "qa",
      "score01": 0.94,
      "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": 1,
          "details": {
            "actual": 50400
          }
        },
        {
          "kind": "rubric",
          "score": 0.7,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.94
      },
      "metrics": {
        "ttftMs": 9516,
        "tokens": 372,
        "tokPerSec": 1675.6756756756756,
        "wallMs": 9738,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.94,
          "wallMs": 9738
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/hyou-csv-02.md"
    },
    {
      "id": "hyou-csv-03",
      "category": "hyou-csv",
      "categoryLabel": "表/CSV処理",
      "type": "qa",
      "score01": 1,
      "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": 1,
          "details": {
            "actual": 531780
          }
        },
        {
          "kind": "rubric",
          "score": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": 22995,
        "tokens": 2454,
        "tokPerSec": 1120.0365130077591,
        "wallMs": 25186,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 25186
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/hyou-csv-03.md"
    },
    {
      "id": "hyou-csv-04",
      "category": "hyou-csv",
      "categoryLabel": "表/CSV処理",
      "type": "qa",
      "score01": 0.98,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "R-03",
              "R-07"
            ],
            "hits": [
              "R-03",
              "R-07"
            ]
          }
        },
        {
          "kind": "objective",
          "method": "numeric",
          "score": 1,
          "details": {
            "actual": 846000
          }
        },
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "false"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.9,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.98
      },
      "metrics": {
        "ttftMs": 9195,
        "tokens": 393,
        "tokPerSec": 600,
        "wallMs": 9850,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.98,
          "wallMs": 9850
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/hyou-csv-04.md"
    },
    {
      "id": "keiyaku-01",
      "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": [
              "4"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": 17153,
        "tokens": 1583,
        "tokPerSec": 1046.265697290152,
        "wallMs": 18666,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 18666
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/keiyaku-01.md"
    },
    {
      "id": "keiyaku-02",
      "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": [
              "633600"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": 12637,
        "tokens": 955,
        "tokPerSec": 724.5827010622154,
        "wallMs": 13955,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 13955
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/keiyaku-02.md"
    },
    {
      "id": "keiyaku-03",
      "category": "keiyaku",
      "categoryLabel": "契約・見積照合",
      "type": "qa",
      "score01": 0.2,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "values": [
              "1"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.4,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "reliability": 0.2
      },
      "metrics": {
        "ttftMs": 13353,
        "tokens": 1310,
        "tokPerSec": 384.9544519541581,
        "wallMs": 16756,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.2,
          "wallMs": 16756
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/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": [
              "10,000,000円"
            ],
            "hits": [
              "10,000,000"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": 14987,
        "tokens": 1191,
        "tokPerSec": 1293.159609120521,
        "wallMs": 15908,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 15908
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/keiyaku-04.md"
    },
    {
      "id": "keiyaku-05",
      "category": "keiyaku",
      "categoryLabel": "契約・見積照合",
      "type": "qa",
      "score01": 0.76,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "includes",
          "score": 1,
          "details": {}
        },
        {
          "kind": "rubric",
          "score": 0.6,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.76
      },
      "metrics": {
        "ttftMs": 14022,
        "tokens": 1589,
        "tokPerSec": 329.9418604651163,
        "wallMs": 18838,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.76,
          "wallMs": 18838
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/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": 8999,
        "tokens": 261,
        "tokPerSec": 628.9156626506025,
        "wallMs": 9414,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 9414
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/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": 4396,
        "tokens": 513,
        "tokPerSec": 1573.6196319018404,
        "wallMs": 4722,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 4722
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/longctx-02.md"
    },
    {
      "id": "mail-01",
      "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": 1656,
        "tokens": 548,
        "tokPerSec": 191.40761439049948,
        "wallMs": 4519,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 4519
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/mail-01.md"
    },
    {
      "id": "mail-02",
      "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": 7590,
        "tokens": 419,
        "tokPerSec": 170.95063239494084,
        "wallMs": 10041,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 10041
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/mail-02.md"
    },
    {
      "id": "mail-03",
      "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": 9309,
        "tokens": 480,
        "tokPerSec": 375.88097102584186,
        "wallMs": 10586,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 10586
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/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": 7704,
        "tokens": 504,
        "tokPerSec": 170.1553004726536,
        "wallMs": 10666,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 10666
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/mail-04.md"
    },
    {
      "id": "shisya-rules-01",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 1,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "第5条"
            ],
            "hits": [
              "第5条"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": 14040,
        "tokens": 1286,
        "tokPerSec": 533.3886354209872,
        "wallMs": 16451,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 16451
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/shisya-rules-01.md"
    },
    {
      "id": "shisya-rules-02",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 0,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "error": "answer is not valid JSON",
            "invalidJson": true
          }
        },
        {
          "kind": "rubric",
          "score": 0,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0
      },
      "metrics": {
        "ttftMs": 469467,
        "tokens": 53485,
        "tokPerSec": 4080.332621299969,
        "wallMs": 482575,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0,
          "wallMs": 482575
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/shisya-rules-02.md"
    },
    {
      "id": "shisya-rules-03",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 0.44000000000000006,
      "passed": false,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 0,
          "details": {
            "values": [
              "第22条第3項"
            ],
            "hits": []
          }
        },
        {
          "kind": "rubric",
          "score": 0.8,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.44000000000000006
      },
      "metrics": {
        "ttftMs": 12916,
        "tokens": 1048,
        "tokPerSec": 579.0055248618785,
        "wallMs": 14726,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.44000000000000006,
          "wallMs": 14726
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/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",
              "要申請・要承認",
              "不可",
              "要申請・要承認"
            ],
            "hits": [
              "不可"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.8,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "reliability": 0.88
      },
      "metrics": {
        "ttftMs": 14016,
        "tokens": 1371,
        "tokPerSec": 468.8782489740082,
        "wallMs": 16940,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.88,
          "wallMs": 16940
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/shisya-rules-04.md"
    },
    {
      "id": "shisya-rules-05",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 1,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "14",
              "限度10日、うち有給3日",
              "取得不可",
              "取得可、限度5日、全無給"
            ],
            "hits": [
              "14"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 1,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 1
      },
      "metrics": {
        "ttftMs": 22982,
        "tokens": 2615,
        "tokPerSec": 757.531865585168,
        "wallMs": 26434,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 1,
          "wallMs": 26434
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/shisya-rules-05.md"
    },
    {
      "id": "shisya-rules-06",
      "category": "shisya-rules",
      "categoryLabel": "社内規程RAG",
      "type": "qa",
      "score01": 0.9175,
      "passed": true,
      "layerScores": [
        {
          "kind": "objective",
          "method": "json-field",
          "score": 1,
          "details": {
            "values": [
              "在宅勤務規程 第3条, 第5条",
              "情報セキュリティ規程 第10条, CSIRT運用 連絡細則 細則2",
              "CSIRT運用 連絡細則 細則4, 情報セキュリティ規程 第10条第3項",
              ""
            ],
            "hits": [
              "第3条",
              "細則2",
              "第10条"
            ]
          }
        },
        {
          "kind": "rubric",
          "score": 0.85,
          "details": {
            "judge": "claude",
            "source": "workflow-rejudge"
          }
        }
      ],
      "axes": {
        "quality": 0.9175
      },
      "metrics": {
        "ttftMs": 20855,
        "tokens": 2440,
        "tokPerSec": 561.8236242228874,
        "wallMs": 25198,
        "toolCalls": 0,
        "toolCallInvalid": 0,
        "capHit": false,
        "errored": false,
        "crashed": false
      },
      "repeats": [
        {
          "score01": 0.9175,
          "wallMs": 25198
        }
      ],
      "needsAudit": false,
      "transcriptPath": "2026-06-06/transcripts/gpt-oss_20b/shisya-rules-06.md"
    }
  ],
  "rollup": {
    "byCategory": {
      "code-hozen": {
        "scoreMean": 0.825,
        "passRate": 0.75
      },
      "gijiroku": {
        "scoreMean": 0.8550000000000001,
        "passRate": 0.8
      },
      "hyou-csv": {
        "scoreMean": 0.98,
        "passRate": 1
      },
      "keiyaku": {
        "scoreMean": 0.792,
        "passRate": 0.8
      },
      "longctx": {
        "scoreMean": 1,
        "passRate": 1
      },
      "mail": {
        "scoreMean": 1,
        "passRate": 1
      },
      "shisya-rules": {
        "scoreMean": 0.7062499999999999,
        "passRate": 0.6666666666666666
      }
    },
    "byAxis": {
      "quality": 86,
      "reliability": 96,
      "speed": 83,
      "economy": 95,
      "localValue": 95,
      "setup": 70
    },
    "total100": 89,
    "capEvents": [],
    "tasksCompleted": 30,
    "tasksAttempted": 30
  },
  "rescoredAt": "2026-06-05",
  "judge": "claude (rubric, workflow-rejudge) + objective(substring)"
}