Peter Zhou Mistake Bank

Mistake Bank

Use this subskill when a recognized result must become canonical local data.

Core contract:

• Persist one JSON file per subject under the mistake bank.
• Keep SourceDocument state separate from subject mistake records.
• Keep is_wrong=false records for audit, but exclude them from final stats and practice by default.
• Generate review reports from canonical data; do not persist separate long-lived review tables.
• Allocate ids and write JSON only through scripts/allocate_id.py and scripts/json_store.py; see references/schema.md.
• After every successful scripts/intake_recognition.py ingest, present the returned review_report as the next skill step for user review. Focus the user on needs_review=true, low-confidence answers, ambiguous teacher marks, and crop screenshots. If the user does not review, leave the source status as processed; only apply-review --confirm moves it to confirmed.
• Use scripts/intake_recognition.py review-report to regenerate the same derived report, apply-review for user corrections/confirmation, and repair-crops for script-owned screenshot bbox fixes.