{
  "id": "PW-238-essai02-scoped-feedback-v1",
  "created_utc": "2026-10-19T11:48:17.799461+00:00",
  "question": "Does explicit extraction of scoped lessons and retained principles improve unfamiliar project judgements beyond ordinary access to the identical reason-bearing history, while preserving original-use adequacy?",
  "design": "Three conditions, two fresh sessions each, 20 new synthetic prose cases each. Frozen reference sees 4 original examples; ordinary history and scoped lessons see the identical 8 examples, including 4 additional corrections. Each session has stage1 preparation followed by stage2 test with no intervening feedback.",
  "cases": 20,
  "uses": {
    "EXPLORE": 10,
    "DECIDE": 10
  },
  "replicates_per_arm": 2,
  "arms": [
    "frozen",
    "history",
    "scoped"
  ],
  "participant_model": "unknown until actual runner metadata; record exact exposed metadata or unknown, never infer",
  "budgets": {
    "stage1_max_words": 600,
    "stage2_max_words": 1800,
    "calls_per_session": 2,
    "tools": "Read only assigned packet; no external search or other task files",
    "comparability": "History and scoped arms receive byte-identical examples and test cases, equal declared input context and output limits, same configured model/reasoning if available. Hidden inference compute is neither controlled nor claimed equal."
  },
  "information_boundary": "Participants start without author/evaluator context. Stage1 contains no test cases. Stage2 contains no rubric. Author may design rubric, but cannot act as participant. Test/evaluator frozen before any participant begins. Raw reports are retained unchanged.",
  "hypothesis": "Scoped lessons may help distinguish commitment from exploration while retaining evidence and contradiction constraints. No superiority is assumed; a ceiling or ordinary-history equivalence is a substantive negative result.",
  "primary": "Per-run sound-judgement counts out of20 and separate EXPLORE/DECIDE counts out of10, judged against frozen rubric from blinded outputs. All six reports shown, no best-run selection.",
  "secondary": [
    "exact action agreement",
    "reason score",
    "targeted question score",
    "unnecessary blocks",
    "unsupported acceptance",
    "scope collapse",
    "output words",
    "preparation burden as word count only"
  ],
  "limits": [
    "Authored fictional cases and rubric are not independent professional outcome evidence.",
    "Twenty cases are not a population estimate; two same-model sessions expose output sensitivity, not model diversity or human inter-rater reliability.",
    "Case families are related and partially paired; no independence or significance claim.",
    "The reference has less feedback, intentionally. Only history versus scoped isolates preparation method given the same information.",
    "Explicit wording can activate pre-existing model knowledge; this does not demonstrate weight updates, durable memory or acquisition of previously absent concepts.",
    "One later stage measures use of feedback across cases, not sustained longitudinal improvement.",
    "No model or human time saving is claimed."
  ],
  "stop_rule": "Run exactly two sessions per condition with no prompt/data/rubric tuning after any output. Malformed JSON may be mechanically repaired without semantic edits; retain raw and record repair. Failed session may be retried only for technical failure with record; never select a preferred answer.",
  "freeze": "Hash protocol, all participant packets, withheld rubric, author script and verifier. Freeze excludes subsequently generated raw reports and analysis."
}