{
  "version": "two-stage-retrospective-20261023-v1",
  "design": "Exploratory internal evaluation specified after inspecting the single-model and panel results; no new model calls",
  "sourceHashes": {
    "https://eternalsearch.net/omega/grok-generated-confirmation-20261023/report.json": "7c79ef7647a3ae8ba0a8917ab5ebfc509071e23da7993b57a61424029f180e29",
    "https://eternalsearch.net/omega/grok-generated-confirmation-20261023/report-strict.json": "3291d27d10a0a8631cd3964437856f552c08de5edb3f5596f726adf83c54e922"
  },
  "firstStage": [
    "gpt-6-astra",
    "claude-sonnet-5-5"
  ],
  "secondStage": "Every 3-, 4-, or 5-model subset of the five original predictors; retain at least one yes",
  "cohort": "Fixed all-five complete-case cohort within each response-handling condition; paired first-stage, panel-only, cascade and coin use identical identities",
  "endpoints": {
    "core": "At least two of three judges give article-correspondence score >=3",
    "strict": "At least two of three judges give article-correspondence score >=4"
  },
  "uncertainty": "Descriptive pointwise 95% percentile intervals; paired original-task cluster bootstrap, 2000 draws, seed 20261022; no multiplicity adjustment, no independent validation",
  "workload": "One prediction unit per model and hypothesis, reusing first-stage votes in panels that contain that model; not HTTP calls, tokens, money or elapsed time",
  "controls": "Original recorded coin restricted to identical full-cohort or survivor identities; panel-only and first-stage-only controls",
  "invariants": [
    "Final positives are a subset of gate positives",
    "An OR panel including the reused gate leaves its survivors unchanged"
  ]
}