Close audit follow-ups B + E + F (Annex I split, art. 50(4) §2, art. 50(5)) by abk1969 · Pull Request #5 · abk1969/ai_act_compass

abk1969 · 2026-05-15T14:05:35Z

Summary

Stacks on top of #4. Closes the 3 remaining audit follow-ups not addressed in #3 or #4.

Item B — Annex I split (art. 6(1) cumulative test)

art. 6(1) requires BOTH (i) Annex I harmonisation coverage AND (ii) third-party conformity assessment under that regime. The previous single `annexI: 'oui' | 'non'` allowed false-positive `HAUT_RISQUE_ANNEXE_I` when only (i) was true.

Replace `annexI` field with `annexICoverage` + `annexI3rdPartyCA`.
`computeCategory` triggers Annex-I high-risk only when BOTH are `'oui'`.
Step 4 UI: two-part question; sub-question 2 appears only when coverage is `'oui'`. Picking `'oui'` resets sub-question 2 to `null` so the user must answer it explicitly.
4 new tests cover the truth table (both / coverage-only / CA-only / neither). 3 pre-existing tests using the old field shape updated.

Item E — art. 50(4) §2 human-edit exemption

art. 50(4) second subparagraph exempts AI-generated public-interest text from the labelling obligation when the content is under human review / editorial responsibility. The audit flagged that the previous UI gave no path to declare this — any `'genai_text'` selection auto-triggered RISQUE_LIMITE.

Add `answers.art50TextHumanEdit: 'oui' | 'non' | null`.
`computeCategory`: when `'oui'` is claimed and `'genai_text'` is in `art50`, emit a carve-out justification (`ref: 'art. 50(4) §2 carve-out'`) and remove that trigger from RISQUE_LIMITE contribution. Other art. 50 triggers remain active.
UI: Step 6 surfaces the carve-out sub-question only when `'genai_text'` is selected; sub-question state is cleared when `'genai_text'` is deselected.
5 new tests: standard fires, carve-out removes RISQUE_LIMITE alone, other triggers still produce RISQUE_LIMITE, French label, orphan claim is no-op.

Item F — art. 50(5) accessibility caveat

art. 50(5) requires transparency information to be clear, distinguishable, provided at the latest at first interaction/exposure, and compliant with accessibility requirements under Directive 2019/882/EU (European Accessibility Act).

One new entry in `CHECKLIST.RISQUE_LIMITE` Transparency pillar.
Cites Directive 2019/882/EU explicitly (replaced an initial WCAG reference per code-review feedback — the regulation cites the EAA, not WCAG).

Tests

82 tests pass (+8 across the three items in 4 commits)
Build green, 278 kB / 84 kB gzip

Stacking note

Targets `feat/annex-iii-5-subitems` (#4) as base. When #3 and #4 merge, GitHub will auto-rebase this PR onto main.

Test plan

Step 4: pick coverage "Yes", then 3rd-party CA "Yes" → HAUT_RISQUE_ANNEXE_I
Step 4: pick coverage "Yes", then 3rd-party CA "No" → NOT HAUT_RISQUE_ANNEXE_I
Step 4: pick coverage "No" → advances immediately (3rd-party CA auto-set to "non")
Step 4: pick coverage "No" then change to "Yes" → sub-question 2 is empty (not pre-selected)
Step 6: tick "genai_text", claim human-edit "oui" → verdict is NOT RISQUE_LIMITE (assuming no other triggers); carve-out ref present in justifications
Step 6: tick "genai_text" + "interaction", claim human-edit "oui" → RISQUE_LIMITE via interaction only; carve-out ref present, genai_text ref absent
Step 6: deselect "genai_text" after claiming carve-out → carve-out state cleared in answers
RISQUE_LIMITE verdict checklist: art. 50(5) entry visible in screen + PDF + clipboard report
French UI: art. 50(5) checklist line cites "directive 2019/882/UE"

🤖 Generated with Claude Code

…party CA tests art. 6(1) requires BOTH (i) Annex I harmonisation coverage AND (ii) third-party conformity assessment obligation. The single annexI yes/no allowed false-positive HAUT_RISQUE_ANNEXE_I when only (i) was true. Replace with two cumulative answer fields and a Step 4 two-part question. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ated text When the user declares the AI-generated public-interest text is under human review / editorial control (art. 50(4) second subparagraph), the genai_text trigger no longer contributes to RISQUE_LIMITE. Other art. 50 triggers (chatbot, biocat_emotion, genai_media) remain active. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…MITE checklist art. 50(5) imposes that transparency information be clear, distinguishable, provided at the latest at the first interaction/exposure, and accessible to persons with disabilities. Add the corresponding compliance-deliverable entry. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- Item B: reset annexI3rdPartyCA when coverage flips to 'oui' (avoid stale answer carrying over from a previous 'non' click on coverage) - Item E: add orphan-claim test (art50TextHumanEdit='oui' without genai_text selected is a graceful no-op) - Item F: replace speculative 'WCAG-level accessibility' citation in the art. 50(5) checklist line with the regulation's actual reference (Directive 2019/882/EU European Accessibility Act) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

vercel · 2026-05-15T14:05:41Z

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
ai_act_compass	Ready	Preview, Comment	May 15, 2026 2:05pm

…t tests Closes a TDD gap identified in the audit retrospective: the CHECKLIST and QUICKWINS data structures lived inline in the 2700-line ai-act-compass.jsx and were untested. Items shipped to those structures (notably art. 50(5) accessibility in PR #5) had no regression guard. Changes: - New module src/lib/deliverables.js exports QUICKWINS, CHECKLIST, isFriaItem - ai-act-compass.jsx imports them instead of declaring inline - New test file src/lib/deliverables.test.js with 27 characterization tests: * art. 50(5) entry exists in RISQUE_LIMITE Transparency pillar with EAA cite * Every tier has at least one checklist pillar and one quickwin * All bilingual {en, fr} text shapes are non-empty * art. 4 AI literacy is present in tier-specific checklists for the tiers where personnel-facing training is the primary deliverable * art. 27 FRIA quickwin exists exactly once and ONLY in HAUT_RISQUE_ANNEXE_III * isFriaItem predicate semantics (exact match on 'art. 27', not substrings) Tests: 82 → 109 (+27 in new file). Build: green, no bundle size change (data only relocated). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

abk1969 · 2026-05-16T12:44:45Z

Merged into main via rebase fast-forward (commits applied via stack rebase, see main's history). Branch deleted.

abk1969 and others added 4 commits May 15, 2026 15:34

abk1969 mentioned this pull request May 15, 2026

test: extract CHECKLIST/QUICKWINS to testable module + 27 invariant tests #6

Closed

4 tasks

abk1969 closed this May 16, 2026

abk1969 deleted the feat/audit-followups-b-e-f branch May 16, 2026 12:44

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Close audit follow-ups B + E + F (Annex I split, art. 50(4) §2, art. 50(5))#5

Close audit follow-ups B + E + F (Annex I split, art. 50(4) §2, art. 50(5))#5
abk1969 wants to merge 4 commits into
feat/annex-iii-5-subitemsfrom
feat/audit-followups-b-e-f

abk1969 commented May 15, 2026

Uh oh!

vercel Bot commented May 15, 2026

Uh oh!

abk1969 commented May 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

abk1969 commented May 15, 2026

Summary

Item B — Annex I split (art. 6(1) cumulative test)

Item E — art. 50(4) §2 human-edit exemption

Item F — art. 50(5) accessibility caveat

Tests

Stacking note

Test plan

Uh oh!

vercel Bot commented May 15, 2026

Uh oh!

abk1969 commented May 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant