research.json

{
  "tag": "correctness-calibration-20260729",
  "modifiable_files": [
    "scripts/sweep_trusted_text_classifier.py"
  ],
  "readonly_files": [
    "mib/ingest.py",
    "mib/cli.py",
    "mib/blend.py",
    "mib/decide.py",
    "mib/evidence_cache.py",
    "scripts/sweep_leader_candidates.py",
    "tests/test_blend.py",
    "tests/test_injection.py",
    "tests/test_authority.py",
    "docs/LEADER_TECHNIQUE_EXPERIMENT.md"
  ],
  "run_command": ".venv/bin/python scripts/sweep_trusted_text_classifier.py --cache ../handoff-artifacts/leader-techniques-20260729/train-evidence.jsonl --labels ../mib-doc-challenge/data/train_labels.csv --out /private/tmp/mib-trusted-text-results.json --folds 5 --repeats 5 --splitter stratified",
  "metric_grep": "^research_score:",
  "metric_name": "research_score",
  "metric_direction": "higher",
  "secondary_metrics": [
    "^classification:",
    "^false_approvals:",
    "^safe_repeats:",
    "^finding_changes:"
  ],
  "timeout_seconds": 900,
  "setup_check": ".venv/bin/python -m pytest -q tests/test_blend.py tests/test_injection.py tests/test_authority.py",
  "no_new_dependencies": true,
  "soft_constraints": "Development-only; do not modify runtime files or artifacts. Use only active-case trusted page.text from the frozen cache; remove case ids, filenames, paths, quarantined text, prompt-shaped lines, and exact page numbers. Finding-settled cases must remain deterministic and byte-identical, retain baseline confidence, and never use HGB features. Compare against the exact fold-local 0.6 posterior + 0.4 HGB baseline for non-settled rows. Promote only if combined classification plus calibration improves in at least 4 of 5 paired repeats, candidate false approvals do not exceed baseline in any repeat or in mean, the shuffled-correctness control has exactly 0 of 5 combined wins, and grouped template/damage validation confirms the direction. Predeclared final-fit construction: average the five leakage-safe OOF feature vectors per case, use mean OOF correctness as the soft target, fit one Ridge(alpha=10.0) on exactly 1,000 case-level rows, and use the same aggregation for the grouped audit. A finding-settled training row uses its policy posterior and baseline confidence, zero model probabilities, and settled indicator 1; it never triggers HGB inference. Finding-settled runtime rows bypass calibration and retain baseline confidence. Do not claim OCR gains from cache replay or use full-fit public score as generalization evidence."
}