{
  "schema": "grex-public-experiment-summary/v1",
  "updated": "2026-11-04",
  "scope": "Sanitized aggregate transcription of retained reports. Not raw reproducible data. Private identifiers, operational traces and source bodies are excluded.",
  "runDate": "2026-10-31",
  "sourceRecord": {
    "report": "Can Grex turn documents into a trustworthy dataset?",
    "basis": "Retained study report and final closeout evidence"
  },
  "status": "completed-bounded",
  "design": {
    "familiesPerArm": 10,
    "sameFrozenInputs": true,
    "workerModelsDiffer": true,
    "grexWorker": "MiniMax-M3",
    "directWorker": "Claude Opus 5.5",
    "indeterminateCellsExcludedPerArm": 2
  },
  "results": {
    "direct": {
      "firstPassAccepted": 5,
      "finalAccepted": 5,
      "criticalCells": {
        "correct": 167,
        "missed": 0,
        "wrong": 5
      },
      "aggregateWorkerMinutesFinalApprox": 16
    },
    "grex": {
      "firstPassAccepted": 6,
      "finalAccepted": 6,
      "criticalCells": {
        "correct": 168,
        "missed": 0,
        "wrong": 4
      },
      "aggregateWorkerMinutesFinalApprox": 121,
      "rootRuns": 10,
      "correctionRuns": 9,
      "tokensTotalApprox": 41400000,
      "cachedInputFractionApprox": 0.95
    }
  },
  "limitations": [
    "Small selected corpus; model-confounded comparison.",
    "Interpretive reference defects and unmatched reviewer emphasis.",
    "Operator effort and all-in costs not established.",
    "Later runs have different denominators and scoring revisions."
  ],
  "decision": "Repair contract and evaluate context efficiency separately."
}