Written down, or it did not happen.
Governance here means one thing: a claim about the work can be checked by somebody who was not in the room. Not trusted, not inferred from a confident summary. Checked.
The rules quoted on this page sit in tiers, and the tier is the point. A timeless article changes only by ratification. A standing rule applies to every project unless something explicitly waives it. Project policy binds one body of work. A decision record fixes a choice on the day it was made. Each quote below names its tier, so you can see what it would take to change it.
Three of the four rules below decide what a claim has to be anchored to before it counts. The fourth decides what survives the session, because a standard nobody can find next month is a standard that has already stopped applying. What enforces any of it is a different question, and it has its own page.
Done is a claim about reality.
The most common failure in AI-assisted work is not a wrong answer. It is a confident report of success that nobody checked. Every rule in this group exists to make that impossible to do quietly.
A task is complete only when it is verified against named evidence, not merely declared finished.
A standing rule
Done is only meaningful once it is anchored to something checkable: a test that passed, a reviewer who looked, an output that was compared against a stated expectation.
The same rule, with its commentary
Reported status matches the true state of the work at the time it is reported. On track, done, and blocked are claims about reality, not expressions of intent or hope.
A standing rule
Output produced by an automated party is checked before it is acted on or passed to another system; it carries no more authority than the input that produced it.
A standing rule
In practice this produces things like refusing to accept a green test gate that was run while a development server was still holding the build, and re-measuring a claim on every page of a corpus rather than a sample before allowing a library migration to proceed.
Verification needs something to verify against, so the definition comes before the build and not after the first argument about whether a result is good enough.
02 · What right meansDefine what right means, in a form that can fail.
Accuracy targets are easy to state and impossible to argue with, which is what makes them useless. A definition that cannot be falsified cannot be tested, and a system nobody can test is a system nobody should trust.
A quantity left empty and flagged for review is correct. An invented one is a defect.
An accuracy standard, fixed before any code was written
That single sentence decided the architecture. It made the extraction pipeline responsible for reporting what it could not read, gave the test suite something binary to assert, and gave the client a promise that could be checked against their own completed work rather than taken on faith.
A ceiling of any value is the same defect at a different number. A 380-page set truncates at 150 exactly as it truncates at 5.
A decision record, written the day it was made
The same instinct applied to a page limit: the fix was not a bigger number, it was removing the ceiling, because the number was never the design decision.
Mapped to the frameworks, not certified against them.
Responsible-AI deliverables are mapped to NIST AI RMF, ISO/IEC 42001 and the EU AI Act, with third-party licensing diligence traced to component level rather than package level.
Mapped is the accurate word and it is the one we use. Nothing here is a certification, no auditor has signed any of it, and a page that implied otherwise would be the first thing this site tells you not to trust.
The cheat sheet
You have just been told what these frameworks are mapped to. The next question is which of them lands on you, and when. The glossary and regulation cheat sheet answers it on one page: provider or deployer or neither, what applies and from what date, and what a standard does not buy you. Every entry carries its instrument.
Why component level
A package-level licence check tells you what a dependency declares about itself. A component-level one tells you what it actually contains. The gap between those two is where a licensing obligation hides, in both directions: an exposure you did not know you had, and a restriction you were honouring for no reason.
Read from the instrument, not from a summary of a summary.
Most published guidance on AI regulation is recycled from other guidance, and a surprising amount of it is stale or simply wrong. Everything below is checked against the instrument and dated. When the law moves, this moves.
Re-verified 2026-09-07. This is not legal advice. It tells you where to look; it does not tell you that you comply.
Which one are you?
Get this wrong and every decision downstream is wrong. Most organisations are deployers, and wrongly assume the obligations belong to their vendor.
Develops an AI system or general-purpose model, or has one developed, and places it on the market under its own name.
The catch: You can become one by substantially modifying a system, or by putting your name on someone else’s.
Uses an AI system under its own authority, other than for personal non-professional activity. Most organisations are here.
The catch: Deployer duties are lighter than provider duties. They are not nil, and several US state laws put real obligations here.
A named, accountable individual required by DIFC Regulation 10 to operate a high-risk autonomous system commercially - status and tasks close to a Data Protection Officer.
The catch: Neither the EU nor the US mandates a personal role of this kind. In the Gulf it is a hiring decision, not a policy one.
What applies, and when
| Date | Obligation | Status |
|---|---|---|
| 2 Feb 2025 | EU - prohibited practices, and the Art. 4 AI literacy duty | In force |
| 2 Aug 2025 | EU - the penalty regime, Chapter XII | In force |
| 27 Jul 2026 | Digital Omnibus on AI - Regulation (EU) 2026/1744 enters into force | Amends Art. 4, defers high-risk |
| 2 Aug 2026 | EU - transparency obligations, Art. 50, including Art. 50(2) watermarking | In force. Not deferred. |
| 2 Aug 2026 | EU - national supervision of the Art. 4 literacy duty begins | In force |
| 2 Dec 2026 | EU - two prohibitions added by the Omnibus begin to apply, under Art. 5 | Upcoming |
| 2 Dec 2026 | EU - the four-month transitional for Art. 50(2) ends, for systems placed on the market before 2 Aug 2026 | Upcoming |
| 1 Jan 2027 | Colorado - the replacement framework, covered ADMT | Under federal challenge |
| 2 Dec 2027 | EU - Annex III high-risk obligations | Deferred, not cancelled |
| 2 Aug 2028 | EU - Annex I high-risk obligations, embedded in regulated products | Deferred, not cancelled |
Deferred is not cancelled. Conformity assessment has long lead times. A 2027 date is a 2026 programme.
Standards, and what they do not buy you
| Framework | What it is | What it is not |
|---|---|---|
| ISO/IEC 42001 | An AI management system standard. Certifiable, and genuinely useful for structuring accountability. | Not a route to EU AI Act conformity. The Commission’s position is that it does not qualify for presumption of conformity. |
| ISO/IEC 42005 | AI system impact assessment guidance. | Guidance, not an obligation. |
| ISO/IEC 42006 | Requirements for bodies auditing AI management systems. | Relevant when choosing a certification body, not when building. |
| NIST AI RMF | Govern, Map, Measure, Manage. Voluntary, and the most usable structure for a programme that has to start somewhere. | Voluntary. Not a compliance defence in the EU. |
| Harmonised standards | Only standards cited in the Official Journal confer presumption of conformity, under Art. 40. | Nothing else does, however official it looks. |
| OWASP GenAI / Agentic Top 10 | The practical technical risk list - prompt injection, tool misuse, memory poisoning. | Not a governance framework. It is the engineering half. |
Each of these is circulating right now.
Including in AI-generated summaries, which is how a stale claim outlives the instrument that made it false. Checking one is cheap. Repeating one is expensive.
The Omnibus deferred high-risk obligations and nothing else. Article 50 transparency applied from 2 August 2026 and was not deferred. Article 4 literacy and the prohibitions have applied since February 2025, and the penalty regime since August 2025.
Regulation (EU) 2026/1744 · AI Act Arts. 99, 113
It requires neither. Both came from SB 1047, which was vetoed. SB 53 requires a published frontier AI framework, transparency reports, critical safety incident reporting, and whistleblower protections.
SB 53, Transparency in Frontier AI Act, signed 29 Sep 2025
The Commission states that ISO/IEC 42001:2023 does not qualify for presumption of conformity - its goals and definitions are not aligned with the quality management system the Act requires. Only harmonised standards cited in the Official Journal do that. Certification is still worth having. It is not a conformity route.
European Commission, AI Act standardisation position
Colorado repealed and replaced its AI Act before the original ever took effect, dropping the duty of care, the mandatory risk programme and the impact assessments. The replacement takes effect 1 January 2027 and is itself under federal challenge, with enforcement of the predecessor stayed since April 2026.
SB 26-189, signed 14 May 2026 · xAI v. Weiser, D. Colo.
Softened, not removed and not deferred: from a duty to ensure a sufficient level of AI literacy to a duty to take measures to support the development of one. That is an obligation of effort rather than of result, and it still binds every deployer. What did change is the enforcement - the Omnibus moved it out from under the hard penalty pressure and to national supervision from 2 August 2026.
Regulation (EU) 2026/1744 · AI Act Art. 4
A definition and a standard are only as durable as the record of them. This is the rule the other three depend on, and the one that is easiest to skip on a day when the work is going well.
06 · What survivesWhat is written governs. What was only said evaporates.
The test for a good record is not whether it is thorough. It is whether a stranger could pick up the work tomorrow and continue correctly without asking anyone what happened.
The written record is sufficient for a successor to continue the work without relying on anyone's memory of it.
A standing rule
Record over memory. What is written down governs; what was only said survives exactly as long as someone happens to remember it.
A ratified instrument, and the reasoning filed with it
Every incident produces either a change to the rules or a recorded decision not to change them.
A standing rule
A standing decision spoken in conversation is written to a versioned file the same turn. Machine-local memory is cache, never record.
A timeless article, changed only by ratification
A rule nothing enforces is a preference.
Everything on this page is written down and could be ignored by anyone who chose to. What makes the difference is the part that refuses.