06-reference/research

ai act article 113 operative text check

2026-08-21·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
eu-ai-actregulationdigital-omnibuscompliance-datesprimary-source-verification

The Enacting Article Says 2 December 2027 Flat: No Condition, No Trigger, and Three Things Recital 40 Never Mentioned

Not legal advice. This is a text-verification pass by a non-lawyer agent. It reports what the enacting words say, not what they mean for any particular product.

The question

"Does the operative text of amended AI Act Article 113(3) match recital 40 exactly, and does it contain any conditional trigger on the 2 December 2027 Annex III date?"

Derivative follow-up #1 from [[2026-08-15-eu-digital-omnibus-ai-official-journal-check]], which confirmed the date from recital 40 plus consolidation metadata after two full-text fetches truncated before the enacting article. Recitals are interpretive, not operative, so that confirmation sat one notch below its headline.

Answer, up front. The evidence gap is closed, and the closure is at primary-source tier on two independent documents.

Sub-question Answer Evidence tier
(a) Does operative Art. 113(3) match recital 40 exactly? No, and it was never going to. Substantively consistent on both dates; materially richer in three respects recital 40 is silent on Primary, authentic OJ text read in full
(b) Any conditional trigger on 2 December 2027? None. Unconditional fixed calendar date, "shall apply from" Primary, authentic OJ text read in full
Cross-check on consolidated AI Act Identical wording under the M1 amendment marker Primary, EUR-Lex consolidation 02024R1689-20260727

The operative text, verbatim. Regulation (EU) 2026/1744, Article 1, point (40)(b), replacing point (c) of the third paragraph of Article 113:

"(c) Chapter III, Sections 1, 2, and 3, with the exception of Article 6(5), shall apply from: (i) 2 December 2027 as regards AI systems classified as high-risk pursuant to Article 6(2) and Annex III; and (ii) 2 August 2028 as regards AI systems classified as high-risk pursuant to Article 6(1) and Annex I;"

That is the whole of it. No proviso, no subordinate clause, no cross-reference to a further condition.

What we already know (from the vault)

What the web says

Method note, because the last pass failed here. Both prior attempts used WebFetch, which truncated before the enacting article on documents of this length. This pass fetched with curl to disk (353 KB and 854 KB of HTML respectively), stripped tags locally, and grepped to the target provisions. Full documents, zero truncation, both gates independent. The tool was the constraint, not the source.

Convergences and contradictions

Synthesis for RDCO

The substantive legal answer is small and the process answer is large. Legally: the Annex III clock is a plain calendar date with nothing hanging off it, the Commission cannot move it administratively, and anyone planning against 2 December 2027 can treat it as fixed until the European Parliament and Council pass another regulation. That is a better planning input than a recital-tier date, because a recital-tier date always carries an unquantified tail risk that some proviso downstream reopens it. The tail is now closed. For any future RDCO surface that touches an Annex III use case, the date is a date.

What is worth more than the date is what the operative text held that the recital did not. Recital 40 is an honest and complete account of the Chapter III deferral, and it is a materially incomplete account of the amendment to Article 113. Three live obligations - the Article 6(5) carve-out, the 2 December 2026 prohibitions date, and the 27 July 2026 enforcement-provisions date - exist only in the enacting words. A researcher who stopped at the recital would have gotten the headline right and the compliance surface wrong, and would have had no signal that anything was missing. That is the specific shape of recital-tier risk, and it is not "the recital might be inaccurate." It is "the recital is accurate about a smaller thing than you asked."

The methodological lesson generalizes past EU law. Two prior fetches failed the same way for the same reason, and the failure looked like an inconclusive result rather than a tool limit. WebFetch renders a long document through a summarizing model and returns what fits; on a 350 KB regulation the enacting articles sit past that horizon while the recitals sit inside it, which biases every long-legal-document pass toward interpretive text and away from binding text. Fetching with curl and grepping locally removed the constraint entirely and cost one extra command. This should become the default for any primary-source pass against a document over roughly 100 KB: statutes, filings, standards, long SEC narrative sections, terms of service. The [[2026-05-19-verification-as-independent-worker-pattern]] gate requirement is only as good as the fetch behind it - a gate that structurally cannot reach the primary source is not a gate.

The last thing to bank is that the 15 August brief's caveat behaved exactly as a caveat should. It named the gap, sized it, said what would close it, and produced a follow-up that took one pass to resolve. Two briefs later the vault has a primary-source-verified compliance date and a reusable fetch pattern. The alternative - stamping the recital finding as confirmed - would have produced a note that read stronger and was worth less, and nobody would have gone back.

Why this is in the vault

It discharges follow-up #1 of [[2026-08-15-eu-digital-omnibus-ai-official-journal-check]], upgrading the 2 December 2027 Annex III date from recital-tier to enacting-text-tier and removing the conditional-trigger tail risk that brief explicitly left open. It also establishes the curl-and-grep fetch pattern as the standard for primary-source passes on long documents, after WebFetch truncation caused the same failure twice.

Open follow-ups

Related

Sources

Primary (read in full, no truncation):

Vault: