[abrichr]

Summary

Rewrites the README for a developer who finds openadapt-privacy on GitHub or PyPI, and fixes the small issues the old README described instead of fixing.

README

• The first screen says what the package removes, who it is for, and gives a two-command install with a one-line example.
• Every Python example was run against this branch. tests/test_readme.py checks the quick start, scrub_dict, and recording outputs against the scrubber, so the README can't drift from the code.
• Limits are in one short section: synthetic test data, English only, context-dependent placeholder names (a Luhn-failing card number such as 4111 1111 1111 1112 keeps its first eight digits), labelled street addresses only, OCR misses, and a regex-only CI scan.
• Operator material (fail-closed behavior, the provenance record, review steps, the repository lifecycle link) moved to docs/COMPLIANCE.md, linked once.
• Configuration is now documented through the config object. The old example built a PrivacyConfig that nothing read.
• Adds the Tesseract requirement for screenshot scrubbing, which was missing.
• The scrubbed demo screenshot is regenerated with current code. The old image came from the first release and showed the card number unredacted. The new one shows the real output, including the name in the dark title bar that Tesseract doesn't read; the README says so.
• Image and link URLs are absolute, so they render on PyPI.

Code fixes

• fix: Field labels stay readable: SSN: 923-45-6789 now scrubs to SSN: <US_SSN> instead of <ORGANIZATION>: <US_SSN>, and Email: ... to Email: <EMAIL_ADDRESS>. A label is dropped from the findings only when its whole text is a known label and another finding starts right after it, so the identifier is always still redacted. Labels that are also surnames (Ip, Sin) are not in the list, so Dr. Sin: 555-123-4567 still redacts the name, as 1.1.1 did.
• fix: The street-address recognizer fired only after Home address: or Mail to, so Address: 42 Elm Road passed through unredacted. It now fires after any label ending in address:. The synthetic recall set grows from 26 to 27 cases.
• fix: The SCRUB_FILL_COLOR docstring said the default was blue; it fills red (Pillow reads the integer as 0xBBGGRR).
• fix: openadapt-privacy-scan --help printed maintainer notes from the module docstring. It now describes the scan and its exit codes, and a failed scan prints PHI scan failed: instead of PHI / clinic-path scan failed:.
• CLAUDE.md used a class name that doesn't exist (PresidioScrubber). The PyPI summary is clearer.

Testing

• uv run pytest: 178 passed (Python 3.12, Presidio and en_core_web_sm installed).
• CI's no-Presidio job, reproduced on Python 3.10: 124 passed, 8 skipped.
• Fresh venv: pip install "openadapt-privacy[presidio]" and python -m spacy download en_core_web_sm, then the quick start, using the published 1.1.1.
• openadapt-privacy-scan --self-test and --root on a fresh install.
• uv build, scripts/check_source_boundary.py --require-dist, and twine check: all pass.

Notes

• Because of the fix: commits, merging triggers the semantic-release workflow to publish a patch release, which is also what updates the README on PyPI.
• Not changed: the image path still sends OCR text straight to Presidio, so it doesn't apply the label or UI-imperative filters the text path uses (field labels next to identifiers are still boxed in screenshots), and Tesseract doesn't read light text on dark backgrounds. The OCR miss is listed under Limits; the extra label boxes only over-redact.

🤖 Generated with Claude Code