{
  "audit_id": "d2_sparse_cross_20261028_independent_audit",
  "actor": "/root/protocol_audit",
  "actor_type": "agent",
  "audited_at": "2026-10-28T11:33:55.328027+08:00",
  "status": "passed",
  "scope": "Independent read-only checks of saved rows, scores, selected masks, hashes and aggregates; no real study or full-risk calculation rerun.",
  "audited_artifact_sha256": {
    "rows.jsonl": "31bb125e6abe42169b3701d0637675792c8c302004fba5355024144fcc2bf27e",
    "summary.json": "9da32f6920cb84b22f2fa08fd1ee399f5d9e20ff2528e81069281deaf7ba1c8c",
    "execution.json": "6dce6ad34a5a4fbff71b8c70be5363b35880afc1b59a277449810c3150a44f91",
    "preregistration.json": "4f416705e767e3dd4bcb0ce384f8c2ba93d150b47196f61890dd08584c4a6dd7",
    "input_verification.json": "49c00e7b0b24e426c1d9f9f9739d067e3fd9527df593a863a490b722097d0ff0",
    "numerical_verification.json": "1398ae132ab0c8ee68994656bcb9286f81eb25e62e8f6f49a16807a1f2fd8c5b",
    "basis_manifest.json": "771e8337af8dca0bd2571c75333a4ea86a6094dcf52864ae645d2291b7a5e280",
    "selection_manifest.json": "40003005dfb9df1d14ebf431ff46017c187bb82da68c5c67183a5d9fa33f7f09",
    "focus_case.json": "d091ea12dd8fd5fdec7a8be2fa886873a86e7a02a68dd062d1baf80594a01bb8",
    "focus_curves.npz": "98dc34ab5b783455bf3cb5bfb1f24b10874e86915ad8bdd8ce2088ef9985e4c8"
  },
  "preregistration_timing": {
    "commit": "7dc745d2c4e21acb98ef4fa0a4534c2e5e707eda",
    "commit_unix": 1793158023,
    "execution_started_unix": 1793158027.333351,
    "computation_started_unix": 1793158030.272417,
    "commit_before_execution": true,
    "frozen_sources_match_commit_and_disk": true
  },
  "source_integrity": {
    "d2_manifest_files_rehashed": 3528,
    "prior_run_artifacts_rehashed": 36,
    "basis_files_rehashed": 120,
    "selection_files_rehashed": 360
  },
  "row_checks": {
    "status": "passed",
    "total_rows": 45360,
    "new_evaluations": 38880,
    "methods_including_references": 14,
    "conditions_per_method": 3240,
    "unique_draws": 120,
    "interior_per_method": 3137,
    "boundary_per_method": 103,
    "unique_condition_method_keys": 45360,
    "row_predicates_recomputed": true
  },
  "selection_checks": {
    "status": "passed",
    "task_draws": 360,
    "masks_checked": 4320,
    "top_rankings_independently_resorted": 720,
    "random_masks_reproduced": 3600,
    "all_masks_without_replacement_and_within_strict_five_percent": true,
    "cross_sigma_groups_checked": 4320,
    "noise_levels_per_mask": 9,
    "zero_budget_task_draws": 60,
    "near_degenerate_draws": 21,
    "scope_note": "Mask ranking checked from saved scores; score derivatives were covered by synthetic pre-run checks, not recomputed here from real data."
  },
  "aggregation_checks": {
    "status": "passed",
    "groups": 728,
    "composition": "14 methods × (1 overall + 12 recipes + 3 seeds + 36 recipe-seeds)",
    "all_fields_and_quantiles_recomputed": true,
    "random_distribution_recomputed": true
  },
  "primary_conclusion": {
    "status": "supported_on_development",
    "recipes_passing": 12,
    "required_recipes": 12,
    "change05_time_pass": 3043,
    "time_denominator": 3137,
    "change05_risk_pass": 3200,
    "risk_denominator": 3240,
    "magnitude05_time_pass": 3042,
    "magnitude05_risk_pass": 3199,
    "magnitude05_recipes_passing": 12
  },
  "change_vs_magnitude": {
    "paired_counts": {
      "change_only_time_pass": 7,
      "magnitude_only_time_pass": 6,
      "change_only_risk_pass": 1,
      "magnitude_only_risk_pass": 0,
      "identical_selected_time": 2263
    },
    "sum_mask_overlap": 37876,
    "sum_retained_pairs": 41452,
    "interpretation": "Both pass all 12 recipes and repair the highlighted example. A one-condition aggregate advantage is not persuasive evidence of superiority of change scoring over magnitude scoring; no such superiority was preregistered."
  },
  "focus_case": {
    "condition": {
      "protocol": "d1_biased",
      "task": "sine",
      "n": 32,
      "seed": 6103,
      "sigma2": 1
    },
    "rank": 31,
    "pair_count": 465,
    "retained_pairs": 23,
    "full_time": 255,
    "diagonal_time": 1,
    "change05_time": 255,
    "magnitude05_time": 255,
    "random_times": [
      1,
      1,
      1,
      1,
      1,
      1,
      1,
      1,
      1,
      1
    ],
    "repaired": {
      "full": true,
      "diagonal": false,
      "change05": true,
      "magnitude05": true,
      "random05_00": false,
      "random05_01": false,
      "random05_02": false,
      "random05_03": false,
      "random05_04": false,
      "random05_05": false,
      "random05_06": false,
      "random05_07": false,
      "random05_08": false,
      "random05_09": false
    },
    "breakdown_arithmetic_checked": true,
    "scope": "Previously inspected worst development example, not a blind test."
  },
  "retained_failures": {
    "change05_invalid_bias": 252,
    "change05_invalid_total": 20,
    "change05_harm_above_10pct": 1,
    "worst_change05": {
      "protocol": "d4_homogeneous",
      "task": "additive",
      "n": 24,
      "seed": 6103,
      "sigma2": 0.03125,
      "data": "results/d4_homogeneous_n24_seed6103_data.npz",
      "curve": "results/d4_homogeneous_n24_seed6103_additive_var0.03125.npz",
      "method": "change05",
      "t_reference": 862,
      "t_selected": 32,
      "reference_interior": true,
      "predicted_interior": true,
      "risk_initial": 1.0685294754678578,
      "risk_reference": 0.11845081490743813,
      "full_risk_at_selected": 0.25497398737543964,
      "estimated_risk_at_selected": 0.2279097151727358,
      "regret_raw": 0.1365231724680015,
      "regret_over_initial": 0.12776734344012788,
      "time_factor": 26.9375,
      "time_pass": false,
      "risk_pass": false,
      "min_bias": 0.21659404884202615,
      "min_total": 0.2279097151727358,
      "invalid_bias": false,
      "invalid_total": false,
      "pair_count": 120,
      "retained_pairs": 6,
      "retained_fraction": 0.05,
      "mask_file": "selections/d4_homogeneous_n24_seed6103_additive.npz"
    },
    "negative_bias_and_risk_pass": 252,
    "negative_total_and_risk_pass": 20
  },
  "numerical_verification_saved_record": {
    "status": "verified",
    "saved_curves_checked": 3240,
    "maximum_normalized_reconstruction_error": 7.849543237625767e-11,
    "mask_checks": 4320,
    "maximum_normalized_mask_error": 2.949029909160572e-15,
    "independent_checks": 96,
    "maximum_independent_absolute_error": 2.275957200481571e-15
  },
  "limits": [
    "Development compression only; all condition data and the focus failure were inspected before this round.",
    "No train-only method, causal training intervention, necessity proof, or unseen-domain confirmation.",
    "5% is an upper bound per method over active pairs; low-rank zero-budget conditions remain unchanged.",
    "Passing stopping criteria does not validate every risk curve; negative-risk and worst-regret cases retained.",
    "Ten random masks quantify algorithm randomness, not ten independent datasets.",
    "Near-degenerate eigenspaces retain basis dependence."
  ]
}