Corrections methodology and log

1. Fix the record. The label, evidence map, document or rule is corrected in the current benchmark version. If the change is material to comparability, a new benchmark version is frozen instead and the old one stays immutable.
2. Record the correction. An entry below states what changed, why, the previous and the current result, and the date.
3. Re-run impacted models. Every model whose score depends on the corrected item is re-scored or re-run; the run ids are listed.
4. Update scores and reports. Leaderboard numbers change with the re-run. A published report whose numbers changed gets a new report version; its previous result stays visible on the report page under “Report versioning”.
5. Preserve the audit trail. Old runs, labels and report versions are never deleted.

Report an error: use the case page’s source links to check the record, then write to the address in the private-data notes. Disputed labels are excluded from scoring until resolved.