# Changelog — ai-act-classifier All notable changes to this skill are documented here. Format: `## [vX.Y] — YYYY-MM-DD` --- ## [v1.7] — 2026-07-07 Timeline drift fix + suite routing completion. - `compliance-deadlines.md`: AI regulatory sandbox deadline re-synced with the canonical Omnibus timeline (`../ai-act-high-risk/references/ai-omnibus-timeline-postponements.md`) — **2 Aug 2027** (was stale "2 Aug 2026 (unchanged)"). - NEXT STEPS block now routes high-risk candidates to `/ai-act-high-risk` and Art. 50 cases to `/ai-act-transparency` (previously the suite's newest specialist had no inbound route from the classifier's output block). --- ## [v1.6] — 2026-07-05 Completes the Art. 50 transparency propagation — corrects a residual legal-effect error in `references/art50-transparency.md` (§7.2) that the v1.5 pass missed: - **"Presumption of conformity" removed.** §7.2 stated that Code-of-Practice compliance "creates a presumption of conformity with Art. 50(2) and Art. 50(4) (Art. 56(8))". That mis-applied the **Art. 56** GPAI-Code mechanism to the softer **Art. 50(7)** transparency Code, and asserted an effect the Code's own text disclaims. Reframed to match the `ai-act-transparency` skill's verified position: the Code is voluntary and adherence is **not conclusive evidence** of compliance — a strong evidentiary anchor, not a presumption of conformity or a safe harbour. - **Burden-of-proof bullet corrected** — dropped the "burden of proof shifts to them" framing (a consequence of the presumption model); the ultimate legal test remains Art. 50 itself, which only the CJEU can authoritatively interpret. ## [v1.5] — 2026-07-04 Art. 50 transparency legal-accuracy pass — propagates the `ai-act-transparency` v1.1 corrections into `references/art50-transparency.md`, plus an adjacent Art. 53(2) fix in `references/scope-exclusions.md`: - **Art. 53(1)(d) mis-cite corrected** — 53(1)(d) is the GPAI *training-content summary* duty, not a "policy to comply with Art. 50(2)". Art. 50(2) marking binds the **AI-system** layer (not GPAI-specific); model-level marking is a value-chain best practice (draft Guidelines paras 24/70 *encourage*; CoP Measure 1.1.2 *expects* of signatories) — not an Art. 53 duty. - **Machine translation moved IN scope** — a translation engine is in scope of Art. 50(2) (draft Guidelines para 54); it had been listed OUT. - **200-token text-watermark rule added** — final Code requires imperceptible watermarking of free-form text over 200 tokens; the dropped "provenance certificate" route is noted. - **EU icon set corrected** — three empirically user-tested icons (Basic / Fully AI-Generated / Partially AI-Modified), optional; the "AI" acronym is mandatory, GENERATED/MODIFIED optional and copyright-sensitive. - **EAA over-framing narrowed** — Art. 50(5) requires "applicable accessibility requirements"; assess EAA (Directive (EU) 2019/882) applicability and use WCAG 2.1 AA as the design benchmark. - **Draft→final Code refresh** — re-anchored to the final Code (10 Jun 2026, adequacy assessment pending); status table updated (Guidelines consultation closed 3 Jun 2026; Omnibus adopted 29 Jun 2026, awaiting OJ). - **Art. 53(2) open-source exemption corrected** — the carve-out covers Art. 53(1)(a)+(b) only; the copyright policy (c) and training-content summary (d) are **not** exempted. ## [v1.4] — 2026-06-21 Art. 50 effective-date correction — completes v1.3 (one missed instance). - `art50-transparency.md`: the document header still read "Applicable since: 2 August 2025 (Art. 113(b))". The v1.3 fix corrected the §7.1 status-table row but missed this near-identical twin string in the header. Corrected to **applies from 2 August 2026** (Art. 113 general application; Art. 50 is Chapter IV — not the 113(b) 2 Aug 2025 tranche; legacy 50(2) marking: 2 Dec 2026), consistent with the rest of the file and the canonical timeline. --- ## [v1.3] — 2026-06-21 Art. 50 transparency effective-date correction (Digital Omnibus — pending OJ publication). - `compliance-deadlines.md`: un-bundled Art. 50 from the 2 Dec 2027 Annex III high-risk phase → **2 Aug 2026** (legacy 50(2) marking: 2 Dec 2026). - `art50-transparency.md`: Art. 50 was mis-attributed to the 2 Aug 2025 / Art. 113(b) tranche — corrected to the Art. 113 general application date of **2 Aug 2026** (Art. 50 is Chapter IV). - Date now sourced from the canonical timeline file. `Last verified` → June 2026. --- ## [v1.2] — 2026-05-19 Routing handoff and AI Omnibus timeline propagation. - Added Step 3.5 routing: high-risk branch now delegates depth assessment to the new ai-act-high-risk skill. - Updated description to flag the routing handoff. - AI Omnibus 2026 postponements applied to compliance-deadlines.md (Annex III: 2 Aug 2026 -> 2 Dec 2027; Annex I: 2 Aug 2027 -> 2 Aug 2028). - enforcement-framework.md timeline narrative updated. - evals/evals.json case 0 expected_output updated to the Omnibus-postponed date. - SKILL.md disclaimer extended with Omnibus note. ## [v1.1] — 2026-05-13 Content-correctness fix on the classification decision tree. - Add explicit callout above the decision tree clarifying that **GPAI assessment (Step 4) runs in parallel with Steps 1–3**, not sequentially after them. A high-risk AI system can simultaneously be a GPAI model with systemic risk; both regimes apply. - Fix right-edge box-pipe alignment on four lines of the ASCII decision tree (Step 4 inner rows + Step 5 "NO → MINIMAL RISK" branch). - No behavior change; corrects a visual implication that GPAI determination was downstream of the risk-tier routing. ## [v1.0] — 2026-05-12 First **reviewed** release. Eval pass via `/skill-creator` confirmed v0.10 content quality with no fixes required. - 3 realistic test cases run with-skill vs. no-skill baseline: clear high-risk employment (TalentMatch CV screener), gray-zone Art. 5(1)(f) emotion recognition (Berlin customer-service SaaS), GPAI fine-tuning edge case (French legal-tech with Llama 3.1 fine-tune) - Result: 100% pass rate (20/20 assertions) vs. 90.5% baseline - **Diagnostic finding (eval-1):** on the gray-zone Art. 5 case, baseline concluded "prohibited" (broad deployment-context reading); skill correctly concluded "not prohibited, high-risk under Annex III Nr. 1(c)" (subject-centric reading per Commission Guidelines). Same case, opposite legal conclusion — clearest evidence of where the skill earns its keep. - See `../../ai-act-classifier-workspace/iteration-1/` for full eval artifacts (test prompts, outputs, grading, benchmark, viewer HTML) ## [v0.10] — 2026-05-08 Promoted off-process improvements that drifted in the CLAUDE_SKILL_EU AI Act/ working folder between 2026-03-09 and 2026-03-15. Status was **audited** as of 2026-03-15 (see `docs/ai-act-suite/2026-03-15-audit.md` — scored 4.6/5 across 12 dimensions). - Description rewritten to "should be used when" pattern with explicit trigger phrases - Phase 1 scope check redesigned: system-description-informed exclusion analysis instead of forcing a 7-letter checklist - Phase 3 prohibited-practice screening: silent internal relevance scoring with single-table presentation - Phase 3 Annex III check: auto-pre-screen before user interaction - Added 4 new reference files: art50-transparency.md, compliance-deadlines.md, enforcement-framework.md, jurisdiction-requirements.md ## [v0.9] — 2026-03-09 - Initial portfolio import (pre-review baseline)