cxr-curatebench-review / README_REVIEW.md
cxr-curatebench's picture
Upload reviewer-safe CXR-CurateBench artifact
ab2292f verified
|
Raw
History Blame Contribute Delete
1.53 kB
# Reviewer Path
Start here:
1. Run `python3 scripts/smoke_test.py`.
2. Run `python3 scripts/scan_identity.py`.
3. Run `python3 scripts/score_contract.py examples/toy_predictions.jsonl examples/toy_targets.jsonl` to inspect answer parsing, box parsing, IoU, and Joint@0.5 mechanics on redistribution-safe toy rows.
4. Run `python3 scripts/recompute_bootstrap.py` and `python3 scripts/recompute_failure_taxonomy.py` to summarize reviewer-safe validation exports.
5. Open `croissant.json` and `metadata/rai_metadata.json`.
6. Inspect the frozen contract in `eval/frozen_downstream_eval_20260322/contract.json`.
7. Inspect paper-facing table exports under `eval/frozen_downstream_eval_20260322/analysis_20260326/paper_tables/`.
8. Inspect paired uncertainty in `eval/frozen_downstream_eval_20260322/analysis_20260326/bootstrap_report.md` and `pairwise_branch_comparisons.csv`.
9. Inspect tradeoff/failure summaries in `analysis/failure_taxonomy_primary_residual.csv`.
10. Inspect endpoint harshness diagnostics in `analysis/endpoint_diagnostics/`.
The core branch comparison should be read as a within-contract measurement problem. Joint@0.5 is an intentionally strict exact-answer/single-box endpoint; it supports paired comparison under a frozen contract, not a calibrated clinical endpoint.
The primary residual is source-ID-disjoint from the visible-rationale pool but is not verified at patient/study level. The supplemental VinDr diagnostic is not training-disjoint and is included only as a source-coverage diagnostic.