What problem does it solve?
This Skill provides a repeatable workflow for correcting and reapplying document redactions when initial OCR or box proposals were incomplete or inaccurate, while minimizing the risk of text-layer leaks.
Core Features & Use Cases
- Two-pass review workflow (text first, optional visual QA): run Pass 1 OCR/CSV/text checks first, then only run Pass 2 VLM on targeted flagged pages when needed.
- CSV-driven edits with safety guardrails: edit the existing *_review_file.csv to add/remove/modify normalized bounding boxes and ensure coordinates are valid (0–1) before apply.
- Coverage-first verification before committing changes: enforce use of verify_redaction_coverage so you can iterate until policy satisfaction (pass_strict) and confirm no text-layer leaks on *_redacted.pdf.
- One consistent apply path: apply redactions only via /review_apply using the original PDF plus the matching review CSV, avoiding unsafe custom PDF rewriting.
Quick Start
Use the doc-redaction-modifications Skill to merge your OCR-derived review edits into *_review_file.csv, run the coverage report until pass_strict is true, then call /review_apply once to generate the corrected *_redacted.pdf for verification.
