{
  "schema": "grex-public-experiment-summary/v1",
  "updated": "2026-11-04",
  "scope": "Sanitized aggregate evidence summary. This is a transcription of retained reports, not a raw-data release or independent reproduction. Private task answers, source bodies, internal identifiers and operational traces are excluded.",
  "runDate": "2026-10-26",
  "status": "completed-adoption-held",
  "sourceRecord": {
    "report": "Bounded instruction-refinement pilot result",
    "reportDate": "2026-10-26",
    "evaluation": "Independent result acceptance reconciled outputs, blinded verdicts, accounting and hold disposition",
    "acceptanceScope": "Integrity of comparison and accounting; not candidate quality or production adoption"
  },
  "upstream": {
    "name": "Harness²",
    "revision": "1cfe254e8d25b15f37048c3730134686abe60335"
  },
  "design": {
    "heldoutGroups": 1,
    "sourcePagesInGroup": 3,
    "heldoutAttempts": 4,
    "workerModel": "MiniMax-M3",
    "synthesisAndImprover": "Haiku 5.5",
    "evaluator": "Opus 5.5",
    "execution": "External controller using pinned Grex research implementation",
    "frozenCandidateBeforeHeldout": true,
    "retrySelection": "First independent acceptance; no prior output or feedback supplied to retries",
    "matchedActualComputeEstablished": false,
    "candidateSurface": [
      "worker instructions",
      "tool guidance"
    ]
  },
  "results": {
    "unchanged": {
      "attempts": 1,
      "selectedAccepted": true,
      "allMechanicalChecksPassed": true,
      "selectedUnsupportedClaims": 0,
      "elapsedSeconds": 227.414,
      "heldoutAnthropicAPIListPriceEstimateUSD": 0.160377
    },
    "retries": {
      "attempts": 2,
      "selectedAttempt": 2,
      "selectedAccepted": true,
      "allMechanicalChecksPassed": true,
      "selectedUnsupportedClaims": 0,
      "firstAttemptUnsupportedClaims": 2,
      "elapsedSeconds": 590.436,
      "heldoutAnthropicAPIListPriceEstimateUSD": 0.360492
    },
    "refined": {
      "attempts": 1,
      "selectedAccepted": false,
      "allMechanicalChecksPassed": true,
      "selectedUnsupportedClaims": 2,
      "elapsedSeconds": 213.799,
      "heldoutAnthropicAPIListPriceEstimateUSD": 0.181287
    }
  },
  "cost": {
    "developmentAnthropicAPIListPriceEstimateUSD": 0.381121,
    "heldoutAnthropicAPIListPriceEstimateUSD": 0.702156,
    "totalAnthropicAPIListPriceEstimateUSD": 1.083277,
    "approvedAPICeilingUSD": 10,
    "actualInvoiceOrCreditDebitObserved": false,
    "allInCostUSD": null,
    "unmeasuredEarlierNativeInvocations": 3
  },
  "decision": {
    "productionAdopted": false,
    "productionAdoption": "Hold",
    "additionalStudyExpansion": "Stopped from this pilot",
    "jobBenchRun": false
  },
  "limits": [
    "One convenience source-window group, not three independent tasks or a success-rate estimate.",
    "Source bodies retrieved in October with August metadata, not verified August archival bytes.",
    "Development adapter defects gave the composer noisy feedback; not a clean estimate of upstream method efficacy.",
    "Shared declared resource ceilings, not equal actual compute or realized cost.",
    "Model judging is fallible; no customer-value or general transfer result."
  ]
}