{
 "generated": "2026-09-29",
 "source": "forensic-rca-calibration-study (private), research/opus55_agent; every endpoint recomputed from judged cells by gen_numbers.py and cross-checked against ablation/endpoints.json; provenance block lists source and key per value",
 "design": {
  "cases": 47,
  "arms": 2,
  "repeats": 2,
  "cells_total": 188,
  "cells_per_arm": 94,
  "judge_passes": 3,
  "judge_models": 2,
  "judge_calls": 1128,
  "failed_cells": 0,
  "noninferiority_margin_pp": 10,
  "confidently_wrong_bound_pp": 5,
  "bootstrap_resamples": 10000,
  "ci_level_primary_pct": 90
 },
 "endpoints": {
  "panel": {
   "useful": {
    "opus46": 0.8298,
    "opus55": 0.9468,
    "diff": 0.117,
    "ci90": [
     0.0532,
     0.1915
    ],
    "ci95": [
     0.0426,
     0.2021
    ],
    "cases_opus55_better": 8,
    "cases_opus46_better": 0
   },
   "exact": {
    "opus46": 0.4787,
    "opus55": 0.5957,
    "diff": 0.117,
    "ci90": [
     0.0,
     0.234
    ],
    "ci95": [
     -0.0213,
     0.266
    ],
    "cases_opus55_better": 14,
    "cases_opus46_better": 7
   },
   "confidently_wrong": {
    "opus46": 0.0638,
    "opus55": 0.0,
    "diff": -0.0638,
    "ci90": [
     -0.117,
     -0.0106
    ],
    "ci95": [
     -0.1383,
     -0.0106
    ],
    "cases_opus55_better": 0,
    "cases_opus46_better": 4
   },
   "grounded": {
    "opus46": 0.4894,
    "opus55": 0.1064,
    "diff": -0.383,
    "ci90": [
     -0.4894,
     -0.2766
    ],
    "ci95": [
     -0.5106,
     -0.266
    ],
    "cases_opus55_better": 0,
    "cases_opus46_better": 22
   },
   "primary_pass": true,
   "co_primary_pass": true,
   "exact_symmetric": {
    "opus46": 0.2872,
    "opus55": 0.3298,
    "diff": 0.0426,
    "ci95": [
     -0.0745,
     0.1489
    ]
   },
   "useful_first_repeat_only": {
    "diff": 0.1064,
    "ci90": [
     0.0426,
     0.1915
    ]
   }
  },
  "gemini_only": {
   "useful": {
    "opus46": 0.8404,
    "opus55": 0.9468,
    "diff": 0.1064,
    "ci90": [
     0.0532,
     0.1702
    ],
    "ci95": [
     0.0426,
     0.1809
    ],
    "cases_opus55_better": 8,
    "cases_opus46_better": 0
   },
   "exact": {
    "opus46": 0.4681,
    "opus55": 0.5957,
    "diff": 0.1277,
    "ci90": [
     0.0106,
     0.2447
    ],
    "ci95": [
     -0.0106,
     0.266
    ],
    "cases_opus55_better": 14,
    "cases_opus46_better": 7
   },
   "confidently_wrong": {
    "opus46": 0.0638,
    "opus55": 0.0,
    "diff": -0.0638,
    "ci90": [
     -0.117,
     -0.0106
    ],
    "ci95": [
     -0.1383,
     -0.0106
    ],
    "cases_opus55_better": 0,
    "cases_opus46_better": 4
   },
   "grounded": {
    "opus46": 0.4894,
    "opus55": 0.1064,
    "diff": -0.383,
    "ci90": [
     -0.4894,
     -0.2766
    ],
    "ci95": [
     -0.5106,
     -0.266
    ],
    "cases_opus55_better": 0,
    "cases_opus46_better": 22
   },
   "primary_pass": true,
   "co_primary_pass": true
  }
 },
 "crosscheck": {
  "max_abs_diff_vs_endpoints_json": 0.0
 },
 "self_preference": {
  "score": {
   "gap_opus55_judge": 0.1383,
   "gap_gemini_judge": 0.234,
   "indicator": -0.0957,
   "indicator_ci90": [
    -0.1915,
    0.0
   ]
  },
  "useful": {
   "gap_opus55_judge": 0.1064,
   "gap_gemini_judge": 0.1064,
   "indicator": 0.0,
   "indicator_ci90": [
    -0.0213,
    0.0213
   ]
  }
 },
 "noise": {
  "panel": {
   "opus46": {
    "useful_agreement": 0.9149,
    "useful_flip_cases": 4
   },
   "opus55": {
    "useful_agreement": 0.9787,
    "useful_flip_cases": 1
   }
  },
  "gemini_only": {
   "opus46": {
    "useful_agreement": 0.8936,
    "useful_flip_cases": 5
   },
   "opus55": {
    "useful_agreement": 0.9787,
    "useful_flip_cases": 1
   }
  },
  "prereg_assumed_flip_rate_pct": 36
 },
 "behaviour": {
  "opus46": {
   "cells": 94,
   "usd_per_cell_mean": 2.3962,
   "turns_mean": 20.2128,
   "tool_calls_mean": 35.2128,
   "elapsed_s_mean": 426.6436,
   "thinking_parts_mean": 0.0,
   "usd_per_cell_median": 2.331,
   "elapsed_min_mean": 7.1,
   "elapsed_s_median": 330.1,
   "usd_total": 225.24,
   "cache_read_share_pooled": 0.8856
  },
  "opus55": {
   "cells": 94,
   "usd_per_cell_mean": 1.5848,
   "turns_mean": 15.7447,
   "tool_calls_mean": 20.234,
   "elapsed_s_mean": 259.8691,
   "thinking_parts_mean": 11.5638,
   "usd_per_cell_median": 1.553,
   "elapsed_min_mean": 4.3,
   "elapsed_s_median": 190.1,
   "usd_total": 148.97,
   "cache_read_share_pooled": 0.8254
  },
  "cost_ratio_opus55_over_opus46": 0.6614,
  "time_ratio_opus55_over_opus46": 0.6091,
  "cases_opus55_cheaper": 47,
  "price_basis": "token-priced at Vertex list price, not billed invoices"
 },
 "spend": {
  "agent_usd": 374.21,
  "judge_usd": 14.91,
  "total_usd": 389.12,
  "cap_usd": 450,
  "stop_before_usd": 440,
  "cost_per_judged_cell_usd": 0.0793
 },
 "confirmation_estimate": {
  "usd_per_case_both_arms_two_repeats_judged": 8.28,
  "usd_for_47_new_cases": 389.12
 },
 "cost_gate": {
  "cells_run_before_stop": 4,
  "cells_planned_original": 94,
  "spent_on_those_cells_usd": 8.0,
  "projection_usd": 216.96,
  "threshold_usd": 170,
  "fired": true,
  "naive_projection_usd": 194.91,
  "repeats_after_amendment": 2,
  "checkpoints_after_amendment": 10
 },
 "raw_vs_final": {
  "opus46": {
   "cells": 94,
   "raw_grounded_claims": 91,
   "final_grounded_chips": 46,
   "link_gate_downgraded": 40,
   "zero_link_cells": 4,
   "no_root_cause_link_cells": 43,
   "text_unchanged_cells": 94,
   "raw_grounded_share": 0.9681,
   "final_grounded_share": 0.4894,
   "link_downgrade_share_of_cells": 0.4255,
   "downgraded_useful_pct": 77.8
  },
  "opus55": {
   "cells": 94,
   "raw_grounded_claims": 81,
   "final_grounded_chips": 10,
   "link_gate_downgraded": 71,
   "zero_link_cells": 61,
   "no_root_cause_link_cells": 84,
   "text_unchanged_cells": 94,
   "raw_grounded_share": 0.8617,
   "final_grounded_share": 0.1064,
   "link_downgrade_share_of_cells": 0.7553,
   "downgraded_useful_cells": 69,
   "downgraded_useful_pct": 97.2,
   "downgraded_useful_rounded_pct": 97,
   "bundle_evidence_cells": 67,
   "bundle_only_no_analyzer_cells": 54
  },
  "text_unchanged_total": 188,
  "claimed_confidently_wrong": {
   "opus46": 0.1702,
   "opus55": 0.0213,
   "diff": -0.1489,
   "ci90": [
    -0.234,
    -0.0745
   ]
  }
 },
 "fix": {
  "cells_total": 188,
  "cases": 47,
  "grounded": {
   "a55": 0.1064,
   "f55": 0.1809,
   "diff": 0.0745,
   "ci90": [
    -0.0106,
    0.1596
   ]
  },
  "useful": {
   "a55": 0.9468,
   "f55": 0.883,
   "diff": -0.0638,
   "ci90": [
    -0.117,
    -0.0213
   ]
  },
  "confidently_wrong": {
   "a55": 0.0,
   "f55": 0.0106,
   "diff": 0.0106,
   "ci90": [
    0.0,
    0.0319
   ]
  },
  "exact": {
   "a55": 0.6064,
   "f55": 0.5638,
   "diff": -0.0426,
   "ci90": [
    -0.1277,
    0.0426
   ]
  },
  "primary_threshold_diff": 0.25,
  "primary_pass": false,
  "noninferiority_pass": false,
  "raw_grounded_claims": {
   "f55": 37,
   "a55": 81
  },
  "link_retry_fired_cells": 19,
  "f55_cells": 94,
  "usd_per_cell_mean_f55": 1.6013,
  "judge_usd": 14.63,
  "answers_echoing_new_schema_wording": {
   "f55": 45,
   "a55": 4
  },
  "spend_total_usd_including_agent": 165.15,
  "miss_cells_a55": 5,
  "miss_cells_f55": 11,
  "miss_cells_with_retries": 0,
  "retried_cells_total": 31,
  "retried_cells_useful_pct": 100,
  "hedged_answers_useful_cells": 38,
  "hedged_answers_useful_pct": 84.4,
  "hedged_answers_miss_cells": 7,
  "cases_lost_usefulness": 5,
  "cases_gained_usefulness": 0,
  "repeat_agreement_f55_pct": 89,
  "link_retry_cost_per_retried_cell_usd": 0.18,
  "link_retry_extra_turns_mean": 1.42,
  "link_retry_changed_prose_cells": 16,
  "link_retry_saved_grounded_cells": 1,
  "link_retry_still_missing_link_cells": 6,
  "link_retry_hit_downstream_gate_cells": 12,
  "post_link_downgrades_total": 20,
  "differential_gate_downgrades": 11,
  "entity_gate_downgrades_f55": 1,
  "primary_lower_bound_required_pp": 15
 },
 "timeline": {
  "study_head_at_registration": {
   "commit": "e1b5f58",
   "author_time": "2026-09-26T12:09:28+01:00",
   "author_time_utc": "2026-09-26 11:09:28Z"
  },
  "prereg_first_commit": {
   "commit": "0984a96",
   "author_time": "2026-09-26T13:06:54+01:00",
   "author_time_utc": "2026-09-26 12:06:54Z",
   "prereg_has_amendment_a1": false,
   "prereg_has_deviation_d1": false
  },
  "results_commit": {
   "commit": "36cb012",
   "author_time": "2026-09-26T23:55:25+01:00",
   "author_time_utc": "2026-09-26 22:55:25Z",
   "prereg_has_amendment_a1": true,
   "prereg_has_deviation_d1": true
  },
  "fix_prereg_and_results_commit": {
   "commit": "af98cc8",
   "author_time": "2026-09-27T15:10:02+01:00",
   "author_time_utc": "2026-09-27 14:10:02Z"
  },
  "fix_prereg_added_in": "af98cc8",
  "first_cell_launch_utc": "2026-09-26T11:44:51Z",
  "first_execution_start_utc": "2026-09-26T11:46:42.708447Z",
  "last_execution_end_utc": "2026-09-26T22:27:08.934259Z",
  "judge_key_created_utc": "2026-09-26T22:28:29+00:00",
  "judge_ledger_final_entry_utc": "2026-09-26T22:52:25+00:00",
  "amendment_a1_stamp_utc_in_text": "2026-09-26 12:12Z",
  "deviation_d1_stamp_utc_in_text": "2026-09-26 14:05Z"
 },
 "shipping": {
  "date_utc": "2026-09-27 15:57Z",
  "opus55_notes_not_grounded_share": 0.8936,
  "opus46_notes_not_grounded_share": 0.5106
 },
 "report01": {
  "claude_grounded_pct": 43,
  "claude_exact_pct": 48
 },
 "gate_history": {
  "added_in_library_date": "2026-06-16",
  "report01_wave2_library_tag": "0.7.19",
  "report01_wave2_library_tag_date": "2026-08-14",
  "wave2_replay_labels_matched": 102,
  "wave2_replay_labels_total": 102,
  "prereg_late_commit_minutes": 22
 },
 "gate_replay": {
  "total_cells_replayed": 818,
  "claude_cells_replayed": 282,
  "gemini_cells_replayed": 536,
  "historical_reproduction_pct": 100,
  "live_confirmation_cells": 24,
  "llm_cost_usd": 0
 },
 "gate_v2": {
  "opus55_grounded_pct": 36.2,
  "opus55_grounded_rounded_pct": 36,
  "opus55_grounded_cells": 34,
  "opus55_precision_pct": 97.1,
  "entity_gate_withdrew_opus55_cells": 28,
  "entity_gate_withdrew_opus46_cells": 17,
  "entity_gate_opus46_useful_cells": 15,
  "stream_false_positive_cells": 34,
  "downstream_hits": 20,
  "upstream_hits": 8,
  "secure_stream_hits": 6,
  "wildcard_pid_false_positive_cells": 28,
  "opus46_cw_before_pct": 6.4,
  "opus46_cw_after_pct": 12.8,
  "opus46_cw_delta_pp": 6.4,
  "gemini37_cw_delta_pp": 5.4,
  "gemini38_cw_delta_pp": 3.3
 },
 "case_review": {
  "cases_audited": 4,
  "broken_key_runs_right": 9,
  "rejudged_cells": 11,
  "rejudge_cost_usd": 0.84
 },
 "gate_2tier": {
  "opus55_grounded_cells": 10,
  "opus55_supported_cells": 24,
  "opus55_supported_pct": 25.5,
  "opus55_supported_rounded_pct": 26,
  "opus55_inconclusive_cells": 60,
  "opus55_supported_useful_pct": 100
 },
 "report02_bridge": {
  "total_tickets": 89,
  "escalated_tickets": 42,
  "max_tokens_before": 4096,
  "max_tokens_after": 16000,
  "prod_draws_truncated": 19,
  "opus55_support_closed_exact_pct": 87,
  "opus55_support_closed_usable_pct": 96,
  "opus55_support_closed_grounded_pct": 22,
  "opus55_support_closed_supported_pct": 38,
  "opus55_support_closed_inconclusive_pct": 37
 },
 "literature": {
  "sculley_year": 2015,
  "efron_year": 1993
 },
 "excluded_from_release": [
  "per-case tables (ABLATION_RESULTS.md end, case_table.md, cells_table.md): 47 case ids",
  "case ids 17xxx to 19xxx anywhere in text, alt text or figures",
  "registry path, image digest, Cloud Build and Cloud Run execution ids, job names, project ids",
  "private repository pull request numbers and commit hashes of private repos",
  "judge rationales and abl_calls.jsonl (quote case content)",
  "per-cell and per-case dollar figures (ratios and arm means only, with the list-price basis)",
  "internal gate step names and the exact downgrade predicate (Pedro decides the detail level: DECISIONS.md item 2)",
  "precision of raw grounded claims (97.5% useful, n=81; post hoc, not a calibration measure; kept out of calibration framing)",
  "grounded_v2 tiers and gate_v2 results (later work, planned as a separate report)",
  "symmetric tie-break useful and per-judge gemini-only exact values beyond the null statement",
  "Vertex list prices per million tokens (dated fact, not needed for the argument)",
  "hidden-thinking token estimate (estimator floor is unvalidated)",
  "F55 retry mechanics beyond the aggregate counts (validator wording is unreleased product detail)",
  "smoke run (SMOKE_PREREG.md): feasibility only, no endpoints"
 ],
 "provenance": {
  "design.cases": {
   "source": "research/opus55_agent/ablation/judge/key.json",
   "key": "distinct case_id in cells, exp3_cases.txt order",
   "recomputed": true
  },
  "design.arms": {
   "source": "research/opus55_agent/ablation/judge/key.json",
   "key": "pseudonym_to_arm (A45, A55; r2 = second repeat)",
   "recomputed": true
  },
  "design.repeats": {
   "source": "research/opus55_agent/ABLATION_PREREG.md",
   "key": "AMENDMENT A1 item 1",
   "recomputed": false
  },
  "design.cells_total": {
   "source": "research/opus55_agent/ablation/judge/key.json",
   "key": "len(cells)",
   "recomputed": true
  },
  "design.cells_per_arm": {
   "source": "research/opus55_agent/ablation/judge/key.json",
   "key": "len(cells)/2",
   "recomputed": true
  },
  "design.judge_passes": {
   "source": "research/opus55_agent/ablation/judge/key.json",
   "key": "passes",
   "recomputed": true
  },
  "design.judge_models": {
   "source": "research/opus55_agent/ablation/judge/key.json",
   "key": "panel",
   "recomputed": true
  },
  "design.judge_calls": {
   "source": "research/opus55_agent/ablation/judge/abl_calls.jsonl",
   "key": "line count",
   "recomputed": true
  },
  "design.failed_cells": {
   "source": "research/opus55_agent/ablation/judge/key.json",
   "key": "failed_cells (empty)",
   "recomputed": true
  },
  "design.noninferiority_margin_pp": {
   "source": "research/opus55_agent/ABLATION_PREREG.md",
   "key": "Endpoints: PRIMARY margin 10pp",
   "recomputed": false
  },
  "design.confidently_wrong_bound_pp": {
   "source": "research/opus55_agent/ABLATION_PREREG.md",
   "key": "Endpoints: CO-PRIMARY upper < +0.05",
   "recomputed": false
  },
  "design.bootstrap_resamples": {
   "source": "research/opus55_agent/ablation/abl_analyze.py",
   "key": "N_BOOT via figdata",
   "recomputed": false
  },
  "design.ci_level_primary_pct": {
   "source": "research/opus55_agent/ABLATION_PREREG.md",
   "key": "Endpoints: 90% CI",
   "recomputed": false
  },
  "endpoints.panel.useful.opus46": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.useful.opus55": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.useful.diff": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.useful.ci90": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.useful.ci95": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.useful.cases_opus55_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.useful.cases_opus46_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.exact.opus46": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.exact.opus55": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.exact.diff": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.exact.ci90": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.exact.ci95": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.exact.cases_opus55_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.exact.cases_opus46_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.confidently_wrong.opus46": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.confidently_wrong.opus55": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.confidently_wrong.diff": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.confidently_wrong.ci90": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.confidently_wrong.ci95": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.confidently_wrong.cases_opus55_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.confidently_wrong.cases_opus46_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.grounded.opus46": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.grounded.opus55": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.grounded.diff": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.grounded.ci90": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.grounded.ci95": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.grounded.cases_opus55_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.grounded.cases_opus46_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (panel)",
   "recomputed": true
  },
  "endpoints.panel.primary_pass": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "ci90 lower > -0.10",
   "recomputed": true
  },
  "endpoints.panel.co_primary_pass": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "ci90 upper < +0.05",
   "recomputed": true
  },
  "endpoints.gemini_only.useful.opus46": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.useful.opus55": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.useful.diff": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.useful.ci90": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.useful.ci95": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.useful.cases_opus55_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.useful.cases_opus46_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.exact.opus46": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.exact.opus55": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.exact.diff": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.exact.ci90": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.exact.ci95": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.exact.cases_opus55_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.exact.cases_opus46_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.confidently_wrong.opus46": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.confidently_wrong.opus55": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.confidently_wrong.diff": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.confidently_wrong.ci90": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.confidently_wrong.ci95": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.confidently_wrong.cases_opus55_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.confidently_wrong.cases_opus46_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.grounded.opus46": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.grounded.opus55": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.grounded.diff": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.grounded.ci90": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.grounded.ci95": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.grounded.cases_opus55_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.grounded.cases_opus46_better": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "per-case mean over 2 repeats, A55 minus A45, paired bootstrap seed 0 (gemini_only)",
   "recomputed": true
  },
  "endpoints.gemini_only.primary_pass": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "ci90 lower > -0.10",
   "recomputed": true
  },
  "endpoints.gemini_only.co_primary_pass": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "ci90 upper < +0.05",
   "recomputed": true
  },
  "crosscheck.max_abs_diff_vs_endpoints_json": {
   "source": "research/opus55_agent/ablation/endpoints.json",
   "key": "max over 8 endpoint blocks of |recomputed - study value| for means, diffs and CI bounds",
   "recomputed": true
  },
  "endpoints.panel.exact_symmetric.opus46": {
   "source": "research/opus55_agent/ablation/endpoints.json",
   "key": "reading_a_PANEL.exact_match_symmetric.A45",
   "recomputed": false
  },
  "endpoints.panel.exact_symmetric.opus55": {
   "source": "research/opus55_agent/ablation/endpoints.json",
   "key": "reading_a_PANEL.exact_match_symmetric.A55",
   "recomputed": false
  },
  "endpoints.panel.exact_symmetric.diff": {
   "source": "research/opus55_agent/ablation/endpoints.json",
   "key": "reading_a_PANEL.exact_match_symmetric.diff",
   "recomputed": false
  },
  "endpoints.panel.exact_symmetric.ci95": {
   "source": "research/opus55_agent/ablation/endpoints.json",
   "key": "reading_a_PANEL.exact_match_symmetric.ci95",
   "recomputed": false
  },
  "endpoints.panel.useful_first_repeat_only.diff": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "rep 1 only",
   "recomputed": true
  },
  "endpoints.panel.useful_first_repeat_only.ci90": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "rep 1 only",
   "recomputed": true
  },
  "self_preference.score.gap_opus55_judge": {
   "source": "research/opus55_agent/ablation/judge/abl_per_judge.json",
   "key": "per_judge medians",
   "recomputed": true
  },
  "self_preference.score.gap_gemini_judge": {
   "source": "research/opus55_agent/ablation/judge/abl_per_judge.json",
   "key": "per_judge medians",
   "recomputed": true
  },
  "self_preference.score.indicator": {
   "source": "research/opus55_agent/ablation/judge/abl_per_judge.json",
   "key": "opus judge gap minus gemini judge gap",
   "recomputed": true
  },
  "self_preference.score.indicator_ci90": {
   "source": "research/opus55_agent/ablation/judge/abl_per_judge.json",
   "key": "paired bootstrap seed 0",
   "recomputed": true
  },
  "self_preference.useful.gap_opus55_judge": {
   "source": "research/opus55_agent/ablation/judge/abl_per_judge.json",
   "key": "per_judge medians",
   "recomputed": true
  },
  "self_preference.useful.gap_gemini_judge": {
   "source": "research/opus55_agent/ablation/judge/abl_per_judge.json",
   "key": "per_judge medians",
   "recomputed": true
  },
  "self_preference.useful.indicator": {
   "source": "research/opus55_agent/ablation/judge/abl_per_judge.json",
   "key": "opus judge gap minus gemini judge gap",
   "recomputed": true
  },
  "self_preference.useful.indicator_ci90": {
   "source": "research/opus55_agent/ablation/judge/abl_per_judge.json",
   "key": "paired bootstrap seed 0",
   "recomputed": true
  },
  "noise.panel.opus46.useful_agreement": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "rep1 vs rep2 on same case",
   "recomputed": true
  },
  "noise.panel.opus46.useful_flip_cases": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "cases where rep1 != rep2",
   "recomputed": true
  },
  "noise.panel.opus55.useful_agreement": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "rep1 vs rep2 on same case",
   "recomputed": true
  },
  "noise.panel.opus55.useful_flip_cases": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "cases where rep1 != rep2",
   "recomputed": true
  },
  "noise.gemini_only.opus46.useful_agreement": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "rep1 vs rep2 on same case",
   "recomputed": true
  },
  "noise.gemini_only.opus46.useful_flip_cases": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "cases where rep1 != rep2",
   "recomputed": true
  },
  "noise.gemini_only.opus55.useful_agreement": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "rep1 vs rep2 on same case",
   "recomputed": true
  },
  "noise.gemini_only.opus55.useful_flip_cases": {
   "source": "research/opus55_agent/ablation/judge/abl_scored.json",
   "key": "cases where rep1 != rep2",
   "recomputed": true
  },
  "noise.prereg_assumed_flip_rate_pct": {
   "source": "research/opus55_agent/ABLATION_PREREG.md",
   "key": "POWER: ~36% of cases flip",
   "recomputed": false
  },
  "behaviour.opus46.cells": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "count of status=complete",
   "recomputed": true
  },
  "behaviour.opus46.usd_per_cell_mean": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "usd",
   "recomputed": true
  },
  "behaviour.opus46.turns_mean": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "generations",
   "recomputed": true
  },
  "behaviour.opus46.tool_calls_mean": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "tool_calls",
   "recomputed": true
  },
  "behaviour.opus46.elapsed_s_mean": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "elapsed_s",
   "recomputed": true
  },
  "behaviour.opus46.thinking_parts_mean": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "thinking_parts",
   "recomputed": true
  },
  "behaviour.opus46.usd_per_cell_median": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "usd",
   "recomputed": true
  },
  "behaviour.opus46.elapsed_min_mean": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "elapsed_s / 60",
   "recomputed": true
  },
  "behaviour.opus46.elapsed_s_median": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "elapsed_s",
   "recomputed": true
  },
  "behaviour.opus46.usd_total": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "usd summed",
   "recomputed": true
  },
  "behaviour.opus46.cache_read_share_pooled": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "sum cache_read / sum tokens_in_total",
   "recomputed": true
  },
  "behaviour.opus55.cells": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "count of status=complete",
   "recomputed": true
  },
  "behaviour.opus55.usd_per_cell_mean": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "usd",
   "recomputed": true
  },
  "behaviour.opus55.turns_mean": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "generations",
   "recomputed": true
  },
  "behaviour.opus55.tool_calls_mean": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "tool_calls",
   "recomputed": true
  },
  "behaviour.opus55.elapsed_s_mean": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "elapsed_s",
   "recomputed": true
  },
  "behaviour.opus55.thinking_parts_mean": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "thinking_parts",
   "recomputed": true
  },
  "behaviour.opus55.usd_per_cell_median": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "usd",
   "recomputed": true
  },
  "behaviour.opus55.elapsed_min_mean": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "elapsed_s / 60",
   "recomputed": true
  },
  "behaviour.opus55.elapsed_s_median": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "elapsed_s",
   "recomputed": true
  },
  "behaviour.opus55.usd_total": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "usd summed",
   "recomputed": true
  },
  "behaviour.opus55.cache_read_share_pooled": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "sum cache_read / sum tokens_in_total",
   "recomputed": true
  },
  "behaviour.cost_ratio_opus55_over_opus46": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "mean usd A55 / mean usd A45",
   "recomputed": true
  },
  "behaviour.time_ratio_opus55_over_opus46": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "mean elapsed_s A55 / A45",
   "recomputed": true
  },
  "behaviour.cases_opus55_cheaper": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "per-case usd summed over 2 repeats",
   "recomputed": true
  },
  "behaviour.price_basis": {
   "source": "research/opus55_agent/ABLATION_PREREG.md",
   "key": "Budget: Cell $ (ledger)",
   "recomputed": false
  },
  "spend.agent_usd": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "usd summed over 188 cells",
   "recomputed": true
  },
  "spend.judge_usd": {
   "source": "research/opus55_agent/ablation/judge/spend_ledger.json",
   "key": "total_usd",
   "recomputed": true
  },
  "spend.total_usd": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "round(agent,2) + round(judge,2); matches ABLATION_RESULTS.md Budget line",
   "recomputed": true
  },
  "confirmation_estimate.usd_per_case_both_arms_two_repeats_judged": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "2 x (mean A45 + mean A55) + 4 x judge cost per cell",
   "recomputed": true
  },
  "confirmation_estimate.usd_for_47_new_cases": {
   "source": "research/opus55_agent/ablation/cells/*.json",
   "key": "per case x 47",
   "recomputed": true
  },
  "spend.cap_usd": {
   "source": "research/opus55_agent/ABLATION_PREREG.md",
   "key": "AMENDMENT A1 item 2",
   "recomputed": false
  },
  "spend.stop_before_usd": {
   "source": "research/opus55_agent/ABLATION_PREREG.md",
   "key": "AMENDMENT A1 item 2",
   "recomputed": false
  },
  "spend.cost_per_judged_cell_usd": {
   "source": "research/opus55_agent/ablation/judge/spend_ledger.json",
   "key": "judge total / cells",
   "recomputed": true
  },
  "cost_gate.cells_run_before_stop": {
   "source": "research/opus55_agent/ablation/ABLATION_RESULTS_1rep_stopped.md",
   "key": "Outcome line",
   "recomputed": false
  },
  "cost_gate.cells_planned_original": {
   "source": "research/opus55_agent/ablation/ABLATION_RESULTS_1rep_stopped.md",
   "key": "Outcome line",
   "recomputed": false
  },
  "cost_gate.spent_on_those_cells_usd": {
   "source": "research/opus55_agent/ablation/projection_S2_1rep.json",
   "key": "spent",
   "recomputed": true
  },
  "cost_gate.projection_usd": {
   "source": "research/opus55_agent/ablation/projection_S2_1rep.json",
   "key": "projection",
   "recomputed": true
  },
  "cost_gate.threshold_usd": {
   "source": "research/opus55_agent/ablation/projection_S2_1rep.json",
   "key": "rule: continue iff projection <= 170",
   "recomputed": true
  },
  "cost_gate.fired": {
   "source": "research/opus55_agent/ablation/projection_S2_1rep.json",
   "key": "CONTINUE",
   "recomputed": true
  },
  "cost_gate.naive_projection_usd": {
   "source": "research/opus55_agent/ablation/projection_S2_1rep.json",
   "key": "naive_rest_total + spent + 7 judge reserve",
   "recomputed": true
  },
  "cost_gate.repeats_after_amendment": {
   "source": "research/opus55_agent/ABLATION_PREREG.md",
   "key": "AMENDMENT A1 item 1",
   "recomputed": false
  },
  "cost_gate.checkpoints_after_amendment": {
   "source": "research/opus55_agent/ablation/plan.json",
   "key": "len(history)",
   "recomputed": true
  },
  "raw_vs_final.opus46.cells": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "count",
   "recomputed": true
  },
  "raw_vs_final.opus46.raw_grounded_claims": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "raw_grounding == grounded",
   "recomputed": true
  },
  "raw_vs_final.opus46.final_grounded_chips": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "grounding == grounded",
   "recomputed": true
  },
  "raw_vs_final.opus46.link_gate_downgraded": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "link_gate_downgrade",
   "recomputed": true
  },
  "raw_vs_final.opus46.zero_link_cells": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "final_links == 0",
   "recomputed": true
  },
  "raw_vs_final.opus46.no_root_cause_link_cells": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "final_root_cause_links == 0",
   "recomputed": true
  },
  "raw_vs_final.opus46.text_unchanged_cells": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "blinded_text_unchanged",
   "recomputed": true
  },
  "raw_vs_final.opus46.raw_grounded_share": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "raw grounded / cells",
   "recomputed": true
  },
  "raw_vs_final.opus46.final_grounded_share": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "final grounded / cells",
   "recomputed": true
  },
  "raw_vs_final.opus46.link_downgrade_share_of_cells": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "downgraded / cells",
   "recomputed": true
  },
  "raw_vs_final.opus55.cells": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "count",
   "recomputed": true
  },
  "raw_vs_final.opus55.raw_grounded_claims": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "raw_grounding == grounded",
   "recomputed": true
  },
  "raw_vs_final.opus55.final_grounded_chips": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "grounding == grounded",
   "recomputed": true
  },
  "raw_vs_final.opus55.link_gate_downgraded": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "link_gate_downgrade",
   "recomputed": true
  },
  "raw_vs_final.opus55.zero_link_cells": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "final_links == 0",
   "recomputed": true
  },
  "raw_vs_final.opus55.no_root_cause_link_cells": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "final_root_cause_links == 0",
   "recomputed": true
  },
  "raw_vs_final.opus55.text_unchanged_cells": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "blinded_text_unchanged",
   "recomputed": true
  },
  "raw_vs_final.opus55.raw_grounded_share": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "raw grounded / cells",
   "recomputed": true
  },
  "raw_vs_final.opus55.final_grounded_share": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "final grounded / cells",
   "recomputed": true
  },
  "raw_vs_final.opus55.link_downgrade_share_of_cells": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "downgraded / cells",
   "recomputed": true
  },
  "raw_vs_final.text_unchanged_total": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "blinded_text_unchanged over 188",
   "recomputed": true
  },
  "raw_vs_final.claimed_confidently_wrong.opus46": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "raw grounded claim and MISS; per-case mean; paired bootstrap seed 0 (post hoc)",
   "recomputed": true
  },
  "raw_vs_final.claimed_confidently_wrong.opus55": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "raw grounded claim and MISS; per-case mean; paired bootstrap seed 0 (post hoc)",
   "recomputed": true
  },
  "raw_vs_final.claimed_confidently_wrong.diff": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "raw grounded claim and MISS; per-case mean; paired bootstrap seed 0 (post hoc)",
   "recomputed": true
  },
  "raw_vs_final.claimed_confidently_wrong.ci90": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "raw grounded claim and MISS; per-case mean; paired bootstrap seed 0 (post hoc)",
   "recomputed": true
  },
  "fix.cells_total": {
   "source": "research/opus55_agent/fix/judge/key.json",
   "key": "len(cells)",
   "recomputed": true
  },
  "fix.cases": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "cases with both arms",
   "recomputed": true
  },
  "fix.grounded.a55": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.grounded.f55": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.grounded.diff": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.grounded.ci90": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.useful.a55": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.useful.f55": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.useful.diff": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.useful.ci90": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.confidently_wrong.a55": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.confidently_wrong.f55": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.confidently_wrong.diff": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.confidently_wrong.ci90": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.exact.a55": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.exact.f55": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.exact.diff": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.exact.ci90": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "per-case mean over 2 repeats, F55 minus A55, paired bootstrap seed 0, panel reading",
   "recomputed": true
  },
  "fix.primary_threshold_diff": {
   "source": "research/opus55_agent/FIX_PREREG.md",
   "key": "PRIMARY grounded diff >= +0.25",
   "recomputed": false
  },
  "fix.primary_pass": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "diff >= 0.25 and ci90 lower > 0.15",
   "recomputed": true
  },
  "fix.noninferiority_pass": {
   "source": "research/opus55_agent/fix/judge/fix_scored.json",
   "key": "ci90 lower > -0.10",
   "recomputed": true
  },
  "fix.raw_grounded_claims.f55": {
   "source": "research/opus55_agent/fix/cells_posthoc/*.json",
   "key": "raw_grounding == grounded",
   "recomputed": true
  },
  "fix.raw_grounded_claims.a55": {
   "source": "research/opus55_agent/fix/cells_posthoc/*.json",
   "key": "raw_grounding == grounded",
   "recomputed": true
  },
  "fix.link_retry_fired_cells": {
   "source": "research/opus55_agent/fix/cells_posthoc/*.json",
   "key": "link_retry_fired",
   "recomputed": true
  },
  "fix.f55_cells": {
   "source": "research/opus55_agent/fix/cells_posthoc/*.json",
   "key": "count",
   "recomputed": true
  },
  "fix.usd_per_cell_mean_f55": {
   "source": "research/opus55_agent/fix/cells/*.json",
   "key": "usd",
   "recomputed": true
  },
  "fix.judge_usd": {
   "source": "research/opus55_agent/fix/judge/spend_ledger.json",
   "key": "total_usd",
   "recomputed": true
  },
  "fix.answers_echoing_new_schema_wording.f55": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "section 1, 'Where the accuracy loss comes from', bullet 4 ('45/94')",
   "recomputed": false
  },
  "fix.answers_echoing_new_schema_wording.a55": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "same bullet ('vs 4/94 A55 answers')",
   "recomputed": false
  },
  "fix.spend_total_usd_including_agent": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "Cost and budget: $165.15 of the $200 cap",
   "recomputed": false
  },
  "timeline": {
   "source": "git log and dumps/run.log, cells/*.json, judge/key.json, judge/spend_ledger.json",
   "key": "see fields",
   "recomputed": true
  },
  "shipping.date_utc": {
   "source": "research/NOTES.md",
   "key": "heading 'SHIPPED: forensic Claude leg opus-4-6 -> opus-5-5' (line 2471)",
   "recomputed": false
  },
  "shipping.opus55_notes_not_grounded_share": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "1 - final grounded share",
   "recomputed": true
  },
  "shipping.opus46_notes_not_grounded_share": {
   "source": "research/opus55_agent/ablation/cells_posthoc/*.json",
   "key": "1 - final grounded share",
   "recomputed": true
  },
  "report01.claude_grounded_pct": {
   "source": "pedrovidigal.com/ai-operations/reports/01-agent-on-the-incident-queue/numbers.json",
   "key": "wave2_public.grounded_pct.claude-opus-4-6",
   "recomputed": false
  },
  "report01.claude_exact_pct": {
   "source": "pedrovidigal.com/ai-operations/reports/01-agent-on-the-incident-queue/numbers.json",
   "key": "wave2_public.exact_pct.claude-opus-4-6",
   "recomputed": false
  },
  "gate_history.added_in_library_date": {
   "source": "hagen-ai git, commit 6d2934f (tag v0.5.1)",
   "key": "commit author date",
   "recomputed": false
  },
  "gate_history.report01_wave2_library_tag": {
   "source": "research/NOTES.md",
   "key": "line 2521: 'W2-37 102/102 (0.7.19)'",
   "recomputed": false
  },
  "gate_history.report01_wave2_library_tag_date": {
   "source": "hagen-ai git, tag v0.7.19",
   "key": "commit author date",
   "recomputed": false
  },
  "gate_history.wave2_replay_labels_matched": {
   "source": "research/NOTES.md",
   "key": "line 2521: own-version replay fidelity W2-37 102/102",
   "recomputed": false
  },
  "gate_history.wave2_replay_labels_total": {
   "source": "research/NOTES.md",
   "key": "line 2521",
   "recomputed": false
  },
  "gate_history.prereg_late_commit_minutes": {
   "source": "git log 0984a96 vs run.log",
   "key": "12:06:54Z minus 11:44:51Z",
   "recomputed": false
  },
  "raw_vs_final.opus55.downgraded_useful_cells": {
   "source": "research/opus55_agent/ablation/posthoc_knowledge.json",
   "key": "69 of 71 link-gate downgraded cells useful",
   "recomputed": false
  },
  "raw_vs_final.opus55.downgraded_useful_pct": {
   "source": "research/opus55_agent/ablation/posthoc_knowledge.json",
   "key": "69 / 71",
   "recomputed": false
  },
  "raw_vs_final.opus55.downgraded_useful_rounded_pct": {
   "source": "research/opus55_agent/ablation/posthoc_knowledge.json",
   "key": "round(69 / 71)",
   "recomputed": false
  },
  "raw_vs_final.opus46.downgraded_useful_pct": {
   "source": "research/opus55_agent/ablation/posthoc_knowledge.json",
   "key": "35 / 45 non-grounded raw-grounded useful",
   "recomputed": false
  },
  "raw_vs_final.opus55.bundle_evidence_cells": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "67/94 A55 cells with verified bundle evidence",
   "recomputed": false
  },
  "raw_vs_final.opus55.bundle_only_no_analyzer_cells": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "54/67 rested on bundle queries alone",
   "recomputed": false
  },
  "fix.miss_cells_a55": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "A55 MISS cells",
   "recomputed": false
  },
  "fix.miss_cells_f55": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "F55 MISS cells",
   "recomputed": false
  },
  "fix.miss_cells_with_retries": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "all 11 MISS cells had 0 retries",
   "recomputed": false
  },
  "fix.retried_cells_total": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "31/31 retried cells were useful",
   "recomputed": false
  },
  "fix.retried_cells_useful_pct": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "31/31 = 100%",
   "recomputed": false
  },
  "fix.hedged_answers_useful_cells": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "38/45 hedged answers useful",
   "recomputed": false
  },
  "fix.hedged_answers_useful_pct": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "38/45 = 84.4%",
   "recomputed": false
  },
  "fix.hedged_answers_miss_cells": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "7 of 11 MISSes in hedged answers",
   "recomputed": false
  },
  "fix.cases_lost_usefulness": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "5 cases lost usefulness, 0 gained",
   "recomputed": false
  },
  "fix.cases_gained_usefulness": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "0 cases gained usefulness",
   "recomputed": false
  },
  "fix.repeat_agreement_f55_pct": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "within-arm repeat agreement 89.4%",
   "recomputed": false
  },
  "fix.link_retry_cost_per_retried_cell_usd": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "+$0.18/retried cell",
   "recomputed": false
  },
  "fix.link_retry_extra_turns_mean": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "+1.42 turns",
   "recomputed": false
  },
  "fix.link_retry_changed_prose_cells": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "16/19 changed blinded prose",
   "recomputed": false
  },
  "fix.link_retry_saved_grounded_cells": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "1/19 stayed grounded",
   "recomputed": false
  },
  "fix.link_retry_still_missing_link_cells": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "6/19 still lacked link",
   "recomputed": false
  },
  "fix.link_retry_hit_downstream_gate_cells": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "12/19 hit downstream gates",
   "recomputed": false
  },
  "fix.post_link_downgrades_total": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "20 post-link downgrades in F55",
   "recomputed": false
  },
  "fix.differential_gate_downgrades": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "11/20 differential gate",
   "recomputed": false
  },
  "fix.entity_gate_downgrades_f55": {
   "source": "research/opus55_agent/FIX_RESULTS.md",
   "key": "1/20 entity gate",
   "recomputed": false
  },
  "fix.primary_lower_bound_required_pp": {
   "source": "research/opus55_agent/FIX_PREREG.md",
   "key": "lower bound > +15pp",
   "recomputed": false
  },
  "gate_replay.total_cells_replayed": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "282 Claude + 536 Gemini cells",
   "recomputed": false
  },
  "gate_replay.claude_cells_replayed": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "188 ablation + 94 F55",
   "recomputed": false
  },
  "gate_replay.gemini_cells_replayed": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "Wave 2/3/4 Gemini cells",
   "recomputed": false
  },
  "gate_replay.historical_reproduction_pct": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "282/282 and 536/536 matched",
   "recomputed": false
  },
  "gate_replay.live_confirmation_cells": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "24/24 live agreement",
   "recomputed": false
  },
  "gate_replay.llm_cost_usd": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "$0 trace replay",
   "recomputed": false
  },
  "gate_v2.opus55_grounded_pct": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "34/94 = 36.2%",
   "recomputed": false
  },
  "gate_v2.opus55_grounded_rounded_pct": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "round(36.2)",
   "recomputed": false
  },
  "gate_v2.opus55_grounded_cells": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "34/94",
   "recomputed": false
  },
  "gate_v2.opus55_precision_pct": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "33/34 useful",
   "recomputed": false
  },
  "gate_v2.entity_gate_withdrew_opus55_cells": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "28/28 useful",
   "recomputed": false
  },
  "gate_v2.entity_gate_withdrew_opus46_cells": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "15/17 useful",
   "recomputed": false
  },
  "gate_v2.entity_gate_opus46_useful_cells": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "15/17 useful",
   "recomputed": false
  },
  "gate_v2.stream_false_positive_cells": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "20 downstream + 8 upstream + 6 secure_stream",
   "recomputed": false
  },
  "gate_v2.downstream_hits": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "20 downstream",
   "recomputed": false
  },
  "gate_v2.upstream_hits": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "8 upstream",
   "recomputed": false
  },
  "gate_v2.secure_stream_hits": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "6 secure_stream",
   "recomputed": false
  },
  "gate_v2.wildcard_pid_false_positive_cells": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "28 <*> PID/duration hits",
   "recomputed": false
  },
  "gate_v2.opus46_cw_before_pct": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "6/94 = 6.4%",
   "recomputed": false
  },
  "gate_v2.opus46_cw_after_pct": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "12/94 = 12.8%",
   "recomputed": false
  },
  "gate_v2.opus46_cw_delta_pp": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "+6.4pp",
   "recomputed": false
  },
  "gate_v2.gemini37_cw_delta_pp": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "+5.4pp",
   "recomputed": false
  },
  "gate_v2.gemini38_cw_delta_pp": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "+3.3pp",
   "recomputed": false
  },
  "case_review.cases_audited": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "4 recurring CW cases",
   "recomputed": false
  },
  "case_review.broken_key_runs_right": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "9/9 model runs right",
   "recomputed": false
  },
  "case_review.rejudged_cells": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "11 cells re-judged blind: 9 MATCH, 1 PARTIAL, 1 MISS",
   "recomputed": false
  },
  "case_review.rejudge_cost_usd": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "$0.84",
   "recomputed": false
  },
  "gate_2tier.opus55_grounded_cells": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "10/94 grounded",
   "recomputed": false
  },
  "gate_2tier.opus55_supported_cells": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "24/94 supported (24/24 useful)",
   "recomputed": false
  },
  "gate_2tier.opus55_supported_pct": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "24/94 = 25.5%",
   "recomputed": false
  },
  "gate_2tier.opus55_supported_rounded_pct": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "round(25.5)",
   "recomputed": false
  },
  "gate_2tier.opus55_inconclusive_cells": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "60/94 inconclusive",
   "recomputed": false
  },
  "gate_2tier.opus55_supported_useful_pct": {
   "source": "DIGEST_2026-09-23_25.md",
   "key": "24/24 = 100%",
   "recomputed": false
  },
  "report02_bridge.total_tickets": {
   "source": "pedrovidigal.com/ai-operations/reports/02-four-models-on-the-queue/numbers.json",
   "key": "47 Support-closed + 42 Escalated",
   "recomputed": false
  },
  "report02_bridge.escalated_tickets": {
   "source": "pedrovidigal.com/ai-operations/reports/02-four-models-on-the-queue/numbers.json",
   "key": "42 Escalated",
   "recomputed": false
  },
  "report02_bridge.max_tokens_before": {
   "source": "PROD-CHECK.md",
   "key": "pydantic-ai default 4,096",
   "recomputed": false
  },
  "report02_bridge.max_tokens_after": {
   "source": "PROD-CHECK.md",
   "key": "PR #90 16,000",
   "recomputed": false
  },
  "report02_bridge.prod_draws_truncated": {
   "source": "PROD-CHECK.md",
   "key": "19/19 Opus 5.5 prod draws hit 4,096",
   "recomputed": false
  },
  "report02_bridge.opus55_support_closed_exact_pct": {
   "source": "02-four-models-on-the-queue/numbers.json",
   "key": "87.2% rounded",
   "recomputed": false
  },
  "report02_bridge.opus55_support_closed_usable_pct": {
   "source": "02-four-models-on-the-queue/numbers.json",
   "key": "95.7% rounded",
   "recomputed": false
  },
  "report02_bridge.opus55_support_closed_grounded_pct": {
   "source": "02-four-models-on-the-queue/numbers.json",
   "key": "22.3% rounded",
   "recomputed": false
  },
  "report02_bridge.opus55_support_closed_supported_pct": {
   "source": "02-four-models-on-the-queue/numbers.json",
   "key": "38.3% rounded",
   "recomputed": false
  },
  "report02_bridge.opus55_support_closed_inconclusive_pct": {
   "source": "02-four-models-on-the-queue/numbers.json",
   "key": "37.2% rounded",
   "recomputed": false
  },
  "literature.sculley_year": {
   "source": "Sculley et al. NeurIPS 2015",
   "key": "publication year",
   "recomputed": false
  },
  "literature.efron_year": {
   "source": "Efron & Tibshirani 1993",
   "key": "publication year",
   "recomputed": false
  },
  "excluded_from_release": {
   "source": "this script",
   "key": "editorial list",
   "recomputed": false
  }
 }
}