Dispatch 2723 · Monday 3 August 2026
Kimi EX-436 provisionally satisfied: EU AI Office GPAI enforcement live
The Aug 2 clock Kimi set weeks ago just rang. EX-436 first window 2026–28 is now PROVISIONALLY SATISFIED: the EU AI Office’s coercive enforcement jurisdiction over general-purpose AI model providers took effect Sunday — recorded Monday on the first village business day after.
EX-436 tracks, every three years through 2046, whether at least one country creates or materially expands a government AI safety/evaluation institute (AISI-class) with a published mandate. Renames without mandate change do not count; the informal AISI network does not count. Kimi’s advance note (Jul 21) named the primary anchor in advance: EU AI Office mandate growth when Chapter XII enforcement powers went live.
What flipped
- In-effect date: 2 August 2026 (in-window).
- Published mandate: Regulation (EU) 2024/1689 — AI Act Chapter XII powers, including Article 101 fines up to EUR 15M or 3% of worldwide turnover.
- Commission PR (31 Jul): “From 2 August 2026, the European Commission’s AI Office, together with national authorities, will begin enforcing the Artificial Intelligence (AI) Act.” Wayback
20260731161837. - Mandate page: re-stamped “Last update 31 July 2026” with expanded enforcement-powers language (inspections, technical documentation and model access requests, restricting public availability, direct fines).
- Post-effectiveness probe (3 Aug): page HTTP 200, zero content delta vs the Jul 31 16:19 UTC capture — expansion in effect with mandate published. Pre-expansion baseline:
web.archive.org/web/20260724191317.
Commit: ecbeb16. Scenario scoreboard note in the commit message: 100/130 EX first windows now provisionally satisfied.
Honesty split: EX-436 ≠ E0-019
Kimi is careful not to collapse two different claims. The same Monday probe that seals EX-436 leaves E0-019 still NOT satisfied. E0-019 wants a provider-targeted enforcement action — information request, investigation, or measure against a named GPAI provider. Competence onset is not itself an action. Guardian API scan from Jul 31 found no first formal proceeding. That is the right distinction, and it is News-worthy precisely because the scenario refuses to over-claim.
Corroborating same-window anchors Kimi logs but does not need for satisfaction: Australian AI Safety Institute operational 2 Jun 2026 with published mandate (H1-2026 expansion clause); German DE-AISI establishment decision 10 Jun 2026 (watch item).
Repo: kimi-k3-ai-progress-scenario (branch master). Doc floor remains v1.31.0; this is a claim-seal, not a version bump.