{
  "version": "b_comparator_sensitivity_v1",
  "created_UTC": "2026-10-22T07:58:31.890316+00:00",
  "exposure": "Post hoc exploratory analysis of already examined DEV17/HOLD68; not new untouched validation.",
  "input_sha256": "b45b34ef5453f3ceb766212a6831c1c980949cf851f2e2dc9e8c72b4937e5299",
  "code_sha256": "2be51442f3065e7edd04e0386336d72138c05e996bc9e62858cdcc2088a7617e",
  "records": 1530,
  "policies": [
    "frozen_contract",
    "whole_probability",
    "schema_EOS",
    "leading_object_probability"
  ],
  "design": "All first attempts; retain every family, contract, source condition, consultation and failure. No refit, model call, threshold tuning, response editing or selection.",
  "policies_defined": {
    "frozen_contract": "Exact existing parser: execution, EOS, whole JSON/fence, duplicate keys, schema, finite q, verbatim quote and verdict enum.",
    "whole_probability": "One complete object or complete outer JSON fence, default JSON key semantics, finite numeric q in[0,1]; no EOS/claim/schema checks.",
    "schema_EOS": "Execution/EOS and whole object/outer fence, default JSON key semantics; jsonschema Draft2020 exact keys and claims types/enums; no note-dependent quotation test.",
    "leading_object_probability": "Decode the first brace-started JSON object after an optional opening fence; suffix text ignored; finite numeric q; no EOS/claim/schema checks."
  },
  "interpretation": "Component-policy sensitivity, not a clinical accuracy contest or a comparison with all existing end-user tools. Registered-contract disagreement is not clinical error.",
  "target": "Same released physician grade1 vs2\u20134; invalid response uses q=.5 for every policy.",
  "units": "Response counts for coverage; consultation for resampling. Settings are repeated observations within each consultation.",
  "reporting": "36 groups per policy (DEV/HOLD\u00d73families\u00d72contracts\u00d73conditions), all policy changes versus frozen, valid-only denominator separate.",
  "uncertainty": "Three HOLD macro Brier policy contrasts, average over18 fixed settings inside each consultation;2000 paired consultation bootstrap percentile95% intervals, seed20261022. Pointwise exploratory intervals, no p-values or confirmatory claims.",
  "closure": "Write protocol before examining comparator outputs. Preserve policy gains and harms. No choose-best policy deployment."
}