fix(compliance): correct HIPAA Safe-Harbor citations + drop anonymisation overclaim (CF-18/CF-17) + verify_output tests (CF-14)#1
Merged
Conversation
…tion overclaim (CF-18/CF-17) + verify_output tests (CF-14)
CF-18 — two factual HIPAA 164.514(b)(2)(i) category errors in the manifest
citations, the kind a reviewing radiologist or DPO would catch:
- verify_output.py: the free-text / quasi-identifier catch-all was
mislabelled "(Q) Any other unique characteristic". (Q) is full-face
photographs; the catch-all is (R) "Any other unique identifying number,
characteristic, or code". Relabelled 11 tags plus the burned-in-pixel
finding to (R).
- DeviceSerialNumber was "(N) Device identifiers". (N) is URLs; device
identifiers and serial numbers are (M). Fixed in verify_output.py and the
regulatory_mapping.py Safe-Harbor example mapping (->M, was ->R).
- Qualified the ENS (RD 311/2022) citation: the tool directly evidences
op.exp.8 (audit log) and mp.info.6 (limpieza de documentos); mp.info.3
(cifrado) is NOT implemented and remains the controller responsibility.
- Gated EU AI Act Art. 10 on applicability: added AI_ACT_APPLICABILITY_NOTE
surfaced as an eu-ai-act manifest disclosure. Art. 10 binds only the
provider of a high-risk Annex III system, not anyone who de-identifies.
- Added CITATIONS_VERIFIED_ON constant; DISCLAIMER now carries the
2026-06-01 re-verification date.
CF-17 — dropped the Recital-26 anonymisation overclaim from the Action.D GDPR
clause. It claimed a schema-preserving dummy renders the substituted data
"no longer personal data". That field-level anonymisation claim contradicts
the global PSEUDONYMOUS (not anonymous) classification: salted-hash remapped
UIDs stay reversible with the separately-held salt, so the dataset never
becomes anonymous. Asserting otherwise is exactly the false-anonymisation
overclaim CNIL SAN-2024-013 (Cegedim) sanctioned. Citation narrowed to
Art. 32(1)(a); summary now states the dummy neutralises the field WITHOUT
anonymising the dataset.
CF-14 — new tests/test_verify_output.py (27 tests): metadata residual path,
SQ recursion, value-cleanliness helpers, the happy pixel-OCR path via a fake
pytesseract (no system tesseract binary needed), multiframe mid-slice OCR,
no-pixel objects, VerificationResult status/conclusive/coverage logic,
sampling, empty/garbage dirs, plus regression guards locking the CF-18/CF-17
corrections. verify_output.py coverage ~74 -> 91 percent.
Full suite 226 passed, ruff + mypy --strict clean, golden completeness proof
exits 0. No public API or manifest-hash-affecting change beyond the added
eu-ai-act disclosure.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: fba2054536
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Comment on lines
+635
to
+637
| "in this manifest BIND ONLY the provider of a high-risk AI system within " | ||
| "the meaning of AI Act Art. 6(2) + Annex III (e.g. AI intended as a " | ||
| "medical device safety component, or otherwise listed in Annex III). If " |
There was a problem hiding this comment.
For providers of AI that is itself a medical device or a safety component covered by Annex I product legislation, high-risk classification comes from AI Act Art. 6(1), not Art. 6(2) + Annex III. This new disclosure says Art. 10 binds only Art. 6(2)/Annex III providers and then tells non-Annex-III users Art. 10 does not apply, so an EU medical-device AI provider can receive a manifest that incorrectly disclaims mandatory data-governance obligations. Please include the Art. 6(1)/Annex I path or remove the medical-device example from the Annex III-only gate.
Useful? React with 👍 / 👎.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Post-v0.6.0 correctness polish on the compliance citations and the independent verifier tests. No public API change; the only manifest-payload change is one added eu-ai-act disclosure.
CF-18 — HIPAA Safe-Harbor citation errors
(Q)(full-face photos) instead of(R)(any other unique identifying number/characteristic/code). Relabelled 11 tags + the burned-in-pixel finding.DeviceSerialNumbermislabelled(N)(URLs) instead of(M)(device identifiers and serial numbers). Fixed inverify_output.pyand theregulatory_mapping.pyexample mapping.AI_ACT_APPLICABILITY_NOTEdisclosure (binds only the high-risk Annex III provider).CITATIONS_VERIFIED_ON(2026-06-01) carried in the DISCLAIMER.CF-17 — drop Recital-26 anonymisation overclaim
Action.D GDPR clause claimed a dummy renders data "no longer personal data". That field-level anonymisation claim contradicts the global PSEUDONYMOUS classification (salted-hash remapped UIDs stay reversible with the withheld salt) and is the exact overclaim CNIL SAN-2024-013 (Cegedim) sanctioned. Narrowed to Art. 32(1)(a).
CF-14 — verify_output tests
New
tests/test_verify_output.py(27 tests): metadata residual path, SQ recursion, cleanliness helpers, happy pixel-OCR path via fake pytesseract, multiframe mid-slice, no-pixel objects, VerificationResult logic, sampling, empty/garbage dirs, plus regression guards locking CF-18/CF-17. Coverage ofverify_output.py~74 -> 91 percent.Verification
Full suite 226 passed, ruff + mypy --strict clean,
examples/verify_golden.pyexits 0.