{
  "schemaVersion": "1.0",
  "updatedAt": "2026-08-24",
  "outcomeValues": [
    "NOT_RECORDED",
    "SUCCESS",
    "FAILURE",
    "TIMEOUT",
    "SAFETY_STOP",
    "OPERATOR_ASSIST"
  ],
  "methodology": {
    "purpose": "保存可重現的 VLA 模型評估方法與逐次證據，不提供通用效能或安全門檻。",
    "missingDataRule": "未知保留空白或明寫 unavailable；NOT_RECORDED 不進入分母。",
    "evidenceRule": "每個已記錄 episode 需有 taskId、seed 與 evidenceRef；介入、停止和逾時不算 SUCCESS。",
    "scopeRule": "模擬、真機與混合評估分開標示；READY_FOR_REVIEW 不是效能或安全通過。"
  },
  "method": {
    "runName": "",
    "modelCheckpoint": "",
    "modelRuntime": "",
    "benchmarkVersion": "",
    "evaluationMode": "SIMULATION",
    "taskSuite": "",
    "environmentVersion": "",
    "hardwareRuntime": "",
    "ioProcessing": "",
    "seedPolicy": "",
    "episodesPlanned": "",
    "episodeLimit": "",
    "successDefinition": "",
    "terminationDefinition": "",
    "evidenceRoot": "",
    "owners": "",
    "limitations": ""
  },
  "assessment": {
    "code": "INCOMPLETE_METHOD",
    "descriptiveSuccessRate": null
  },
  "episodes": [
    {
      "episode": 1,
      "taskId": "",
      "seed": "",
      "outcome": "NOT_RECORDED",
      "elapsedSeconds": "",
      "evidenceRef": "",
      "notes": ""
    }
  ]
}
