AI governance and the AI Act inventory

easkills align --root <repo> --reference nist-ai-rmf-1.0   # AI-risk gaps, named
easkills ai-act-register --root <repo> --as-of <date>      # the inventory + its own gaps
easkills ai-act-register --root <repo> --as-of <date> --out docs/ai-act-register.md
easkills propose --root <repo> --from ai-act --as-of <date>   # FRIA skeletons, Art. 27

--lang pl renders the inventory in Polish; the model's own text is not translated.

Non-negotiables

• This tooling generates; it does not attest. The inventory's structure follows the concepts the AI Act reasons in — risk classes, Art. 3 operator roles, Art. 14 human oversight. Its content comes from the model and nowhere else. No completeness against the Act's registration or documentation duties is claimed and no legal review has happened. Say that out loud every time you hand the document over — the generated file says it in its own header, and that header is not decoration to be trimmed.
• Never classify a system to make a report look calm. aiRiskClass is a fact about what the system does, decided by what the Act says, and it is recorded next to the element where the next reader will find it. Downgrading a risk class to clear AIR002 is worse than the finding — the finding is true.
• A gap is disclosed, not closed by editing. The inventory's last section lists every field the model could not fill, with element ids. Filling a field with a guess is worse than the gap.

Scope: what puts a system in the inventory

One property, on the element — declared, never inferred from a type or from a name that sounds like AI:

properties:
  regulatoryScope: ai-act
  aiRiskClass: high            # prohibited | high | limited | minimal
  aiRole: deployer             # provider | deployer | importer | distributor | gpai-provider | ...
  aiOversight: Credit risk committee reviews every declined application
  provider: Scorewell GmbH     # who made it, when someone else did
  contractRef: ctr-2025-scorewell

An element can be in more than one register's scope: a bought credit-scoring service is DORA's ICT third-party risk and the AI Act's high-risk system, and it declares both as regulatoryScope: ai-act dora — the space-joined combination in alphabetical order, still a closed enum. Any other spelling or order is a schema error, never a silent row-drop; under-inclusion is the failure mode that matters in a regulatory report.

An AI system is not a new element type. It is an application component or a technology service like any other, governed by the same oracle — what is AI-specific is the evidence around it: these properties, the principles that bind it (motivation layer), the standards it claims (ea-standards-base), and the accepted risk on it (ea-dispensation).

The five rules

AIR003 works exactly like REG003: ea-dispensation grants a time-bounded waiver — model-risk acceptance has a natural horizon, so the mandatory expiry fits AI risk unusually well — and it surfaces in the inventory without anyone having to remember to mention it. Do not close a dispensation to clear it, and do not leave it out.

gpai-provider is the role of an organisation that provides the general-purpose AI model a system is built on (Chapter V). Those obligations differ from the system's own, and an inventory most often forgets to say the organisation holds them -- record the role when it is true, even when the system itself is minimal-risk.

Fundamental-rights impact assessments (Art. 27)

propose --from ai-act writes an Assessment skeleton (asm-fria-<element>) into staging for each high-risk system the organisation deploys with no assessment bound to it. Art. 27 does not bind every deployer -- it binds bodies governed by public law, private entities providing public services, and deployers of the Annex III point 5(b) and 5(c) systems (creditworthiness, life and health insurance pricing) -- so the first placeholder asks whether it reaches the organisation at all. If it does not, delete the skeleton and record why as a decision (ea-adr); if it does, the assessment's prose, owner and review date stay human, exactly as for every other proposal, and ea-approve promotes it. A re-run skips the system by name while the skeleton sits in staging.

The AI RMF is a reference overlay

An AI risk framework is a taxonomy, so it uses the mechanism ea-align already describes: NIST AI RMF 1.0 ships in references/ (public domain) as a hash-pinned pack — four functions, nineteen categories, each category's outcome statement verbatim because the RMF publishes no short names. An unmapped category is ALN004, and "this category does not apply to us" is a sentence somebody has to sign. Check the pack's NOTICE first: until its structure is verified against NIST AI 100-1 by a human reading, it is a working draft — a legitimate gap-list generator, not an authority to cite.

ISO/IEC 42001 is licensed and stays in the adopter's repository under the adopter's licence, same mechanism (template/reference/README.md) — and nothing may transcribe it from memory.

Keeping the inventory alive

• Regenerate with --as-of, always. Waiver expiry is date-dependent; an inventory with no date on it cannot be reproduced or compared to the last one.
• Read the gap section first, before the tables.
• A new AI system is an intake event. When the inventory gains a row nobody recognises, the model changed before the governance did; route it through ea-delta-ingest and confirm the risk class against what the system actually does.
• Nothing in scope is a legitimate answer. For an organisation running no AI system the Act reaches, ai-act-register reports that no element is tagged and generates nothing. That is the correct output — and the wrong one for an organisation that has simply not tagged its AI systems yet. Know which of the two you are looking at.

Reporting back

The inventory's headline numbers (systems in scope, risk-class breakdown, third-party providers), then three lists that need a human: the fields the model could not fill with their element ids, the open waivers on high- or limited-risk systems, and any AIR005 — which is never a list item to skim but a meeting to call. Close with the sentence the document carries: this is generated from the model, and whether a system is correctly classified and its obligations met is a decision for the people accountable for it.