EU AI Act

Not legal advice

Provision strings are pointers into the instrument, not interpretations of it. Every statement below can serve as evidence toward an obligation; none of them says a system is compliant, and each one states what it does not show.

The first fully populated catalog, and the one with an obligation already in force that the gateway can evidence directly: Art 50(1), the AI-interaction notice, applies from 2 August 2026. Dates follow the Digital Omnibus on AI (Regulation (EU) 2026/1744), which moved Annex III standalone high-risk duties to 2 December 2027 and Annex I embedded systems to 2 August 2028, added prohibitions from 2 December 2026, and extended Art 50(2) marking to legacy systems from the same day.

Registry id	eu-ai-act
Kind	regulation
Jurisdiction	EU
Instrument	Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744
Catalog versions	2024-08 from 2024-08-01 until 2026-07-27 — Regulation (EU) 2024/1689 as published; 2026-09 from 2026-07-27 (current) — As amended by the Digital Omnibus on AI, Regulation (EU) 2026/1744
Verified against instrument	yes
Tagging key (compliance tags)	eu_ai_act
Roles (frameworks.eu-ai-act.role)	authorised_representative, deployer, distributor, importer, product_manufacturer, provider, provider_and_deployer
Classifications (frameworks.eu-ai-act.risk_tier)	high, limited, minimal, unacceptable
Incident deadline rules	standard: 15 days, death: 10 days, widespread: 2 days
Retention floors	high: 183 days
Obligations	30

The catalog has two versions. 2024-08 (in force until 27 July 2026) carries the dates as originally published; 2026-09 carries the Omnibus dates. A statement is evaluated against the version in force on the action’s date, so a report over an older window keeps the older dates.

Deadline rules

Alternative rules: the incident’s classification picks the rule when it names one, else standard applies.

Rule	Deadline from became_aware_at
standard	15 days
death	10 days
widespread	2 days

Deadlines are computed and shown, never enforced — see incidents.

How applicability is computed

Each obligation is evaluated per AI System from the facts declared on its compliance profile — the eu-ai-act block’s role and risk_tier (written through to the resource group’s eu_ai_act_role / eu_ai_act_risk_tier columns, so nothing that existed before breaks), the neutral-core flags such as interacts_with_natural_persons, and the GPAI classification of the models the system binds (flag models gpai / gpai_systemic_risk in the Asset Register; Art 53–55 attach to the model).

Composite systems. The catalog sets composite classification: a system is evaluated at its effective tier — the highest of its declared risk_tier and the tier of every AI System that declares a contributing dependency on it (composite systems). A fraud screener declared minimal under the Art 6(3) carve-out that feeds a high-risk credit decision is evaluated at high; its declared tier stays visible as a claim. A dependency declared non-contributing (with a sealed separability rationale) keeps its own tier.

The board models “obligation X applies to system Y from date Z”, never “system Y is compliant: true” — a boolean dies at the next omnibus, a dated applicability survives it. Three statuses per (obligation, system):

Status	Meaning
applies	The system’s declared facts put it in scope.
not applicable	A declared fact rules it out (e.g. minimal risk → no high-risk duties).
undetermined	The system has not declared the fact the rule needs. An undeclared role cannot rule an obligation out — the honest answer is “declare it”, shown as a to-do, never a pass.

Obligations the platform cannot evaluate alone (whether a system is embedded in an Annex I product, a legacy placement date, whether a deployer is in Art 27 scope) carry a review note naming the determination you need to record. Art 25 can turn a deployer into a provider on rebranding or substantial modification; a contract drift on the system’s closure is exactly that signal — review the role when one fires.

Old obligation keys (annex3-provider-duties, annex3-deployer-duties, art50-transparency, art72-post-market) still resolve: GET /v1/admin/compliance/obligations/{key} maps them to their successors.

Obligations and their evidence

Each obligation lists the statements it is shown against. The basis is one of recomputed, judged, indicator or declared (what the bases mean); the resolver is from the shared library.

AI literacy

art4-ai-literacy · Art 4 · applies from 2025-02-02

Staff dealing with the operation and use of AI systems must have a sufficient level of AI literacy. Applies to providers and deployers of any AI system. Applies to every system for which the framework is active.

What the gateway cannot show: Organisational duty, evidenced by declaration only.

Statement	Basis	Resolver	Does not show
An AI literacy programme is declared.	declared	profile.field_declared {"path": "frameworks.eu-ai-act.ai_literacy_reference"}	Training is organisational; the gateway cannot observe it.

Prohibited practices

art5-prohibitions · Art 5 · applies from 2025-02-02

Practices classified as unacceptable risk are banned outright. Every system is subject to the prohibitions; a system classified unacceptable may not be operated. Applies to every system for which the framework is active.

Statement	Basis	Resolver	Does not show
Sampled runs show no manipulative or deceptive technique.	judged	judge.no_manipulation	Judged by a model over a disclosed sample; stated with human agreement, never merged with recomputed rows.

Additional prohibited practices (Omnibus)

art5-prohibitions-2026 · Art 5 as amended by Reg. 2026/1744 · applies from 2026-12-02 · introduced in catalog 2026-09

The prohibitions added by the Digital Omnibus on AI apply from 2 December 2026. Applies to every system for which the framework is active.

Statement	Basis	Resolver	Does not show
Sampled runs show no manipulative or deceptive technique.	judged	judge.no_manipulation	Judged by a model over a disclosed sample; stated with human agreement, never merged with recomputed rows.

Risk management system

art9-risk-management · Art 9 · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

A continuous, iterative risk-management process across the system’s lifecycle. Applies to: classification high; roles authorised_representative, distributor, importer, product_manufacturer, provider, provider_and_deployer.

What the gateway cannot show: Organisational duty, evidenced by pointer only.

Statement	Basis	Resolver	Does not show
A current risk assessment is on file.	recomputed	evidence.pointer_on_file {"kind": "risk_assessment"}	A pointer to a document produced outside the platform; its content is not evaluated.

Data and data governance

art10-data-governance · Art 10 · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Training, validation and testing data meet the quality criteria of Art 10. Applies to: classification high; roles authorised_representative, distributor, importer, product_manufacturer, provider, provider_and_deployer.

What the gateway cannot show: Training data is outside the gateway; evidenced by pointer only.

Statement	Basis	Resolver	Does not show
A data-governance record is on file.	recomputed	evidence.pointer_on_file {"kind": "data_governance"}	A pointer to a document produced outside the platform; its content is not evaluated.

Technical documentation

art11-technical-documentation · Art 11, Annex IV · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Technical documentation drawn up before placing on the market and kept current. Applies to: classification high; roles authorised_representative, distributor, importer, product_manufacturer, provider, provider_and_deployer.

Statement	Basis	Resolver	Does not show
A documentation export exists for the active contract version, with its digest sealed.	recomputed	docs.technical_documentation {"framework": "eu-ai-act"}	A skeleton generated from the register and contract; its adequacy is not evaluated.

Record-keeping (automatic logging)

art12-record-keeping · Art 12 · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Automatic recording of events over the system’s lifetime, enabling traceability and post-market monitoring. Applies to: classification high; roles authorised_representative, distributor, importer, product_manufacturer, provider, provider_and_deployer.

Statement	Basis	Resolver	Does not show
Every governed action produced a sealed record, including refusals.	recomputed	records.every_verdict_sealed	Integrity is not completeness: covers actions routed through the gateway only; records sealed late by backfill are reported separately.
The tenant log’s tree heads covering the window were witnessed by an independent service.	recomputed	records.witnessed	A receipt shows inclusion at a tree size, not a witness-observed time; a same-operator witness is self-attested.
Approval chains are closed and epoch boundaries match the contract history.	recomputed	records.chain_complete	Checks the chains the gateway opened; actions that never reached the gateway cannot be counted.

Transparency and information to deployers

art13-transparency-to-deployers · Art 13 · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Instructions for use enabling deployers to interpret output and use the system appropriately. Applies to: classification high; roles authorised_representative, distributor, importer, product_manufacturer, provider, provider_and_deployer.

Statement	Basis	Resolver	Does not show
A documentation export exists for the active contract version, with its digest sealed.	recomputed	docs.technical_documentation {"framework": "eu-ai-act"}	A skeleton generated from the register and contract; its adequacy is not evaluated.
Instructions for use are declared.	declared	profile.field_declared {"path": "frameworks.eu-ai-act.instructions_for_use_reference"}	The instructions’ content is not evaluated.

Human oversight

art14-human-oversight · Art 14 · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Designed so natural persons can oversee, interpret, override and stop the system. Applies to: classification high; roles authorised_representative, distributor, importer, product_manufacturer, provider, provider_and_deployer.

Statement	Basis	Resolver	Does not show
A human can override or stop the system.	recomputed	oversight.override_available	Shows the controls exist and are assigned, not that the assigned people are competent.
Human approval requests were decided by humans (requested, decided, lapsed).	recomputed	oversight.exercised	Records the fact of human disposition, not its quality.
Approval latency and denial rate are tracked for movement.	indicator	oversight.not_rubber_stamp	An indicator of movement, never a threshold; it cannot show a decision was considered.
Sampled approval decisions carry a rationale consistent with the request.	judged	judge.approval_rationale	Judged by a model over a disclosed sample; stated with human agreement, never merged with recomputed rows.

Accuracy, robustness and cybersecurity

art15-accuracy-robustness · Art 15 · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Appropriate accuracy, robustness and cybersecurity, declared in the instructions for use. Applies to: classification high; roles authorised_representative, distributor, importer, product_manufacturer, provider, provider_and_deployer.

Statement	Basis	Resolver	Does not show
A current evaluation report is on file.	recomputed	evidence.pointer_on_file {"kind": "eval_report"}	A pointer to a document produced outside the platform; its content is not evaluated.
A current red-team report is on file.	recomputed	evidence.pointer_on_file {"kind": "red_team_report"}	A pointer to a document produced outside the platform; its content is not evaluated.

Quality management system

art17-quality-management · Art 17 · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

A documented quality management system. Applies to: classification high; roles authorised_representative, distributor, importer, product_manufacturer, provider, provider_and_deployer.

What the gateway cannot show: Organisational duty, evidenced by declaration only.

Statement	Basis	Resolver	Does not show
A quality management system is declared.	declared	profile.field_declared {"path": "frameworks.eu-ai-act.qms_reference"}	Organisational duty; the QMS itself is not evaluated.

Automatically generated logs (provider)

art19-log-retention-provider · Art 19 · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Providers keep the automatically generated logs under their control for at least six months. Applies to: classification high; roles authorised_representative, distributor, importer, product_manufacturer, provider, provider_and_deployer.

Statement	Basis	Resolver	Does not show
Logs are kept for at least six months (183 days).	recomputed	retention.floor {"floor_days": 183}	Record identifiers, statements and tree heads are permanent; bodies follow retention.
Every gap in the retained logs is a declared prune.	recomputed	retention.declared_prunes	—

Classification of a composite system

art6-composite-classification · Art 6; Recital 12; draft Commission high-risk classification guidelines (May 2026) · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Where components’ combined outputs materially influence a high-risk decision they are treated as one system: a contributing component is assessed with it, unless it is genuinely separable and does not contribute to the high-risk purpose. Applies to: classification high.

Review: Rests on the Commission’s DRAFT classification guidelines (consultation closed 23 June 2026); re-check against the final text.

What the gateway cannot show: Covers declared dependency edges; the composition checks report undeclared delegation.

Statement	Basis	Resolver	Does not show
Every dependency that contributes to this system’s decisions is assessed at its tier, or carries a sealed separability rationale.	recomputed	dependencies.contributing_assessed	Counts declared dependency edges only; an undeclared delegation is a composition finding, not a statement input. An external agent’s assessment is evidenced by supplier information, not examined.

Written agreements with suppliers

art25-4-supplier-agreements · Art 25(4) · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Third parties supplying AI systems, tools, services or components used in a high-risk system give the provider, by written agreement, the information, capabilities and technical access it needs to comply. Applies to: classification high; roles authorised_representative, distributor, importer, product_manufacturer, provider, provider_and_deployer.

What the gateway cannot show: Evidenced by pointer to the agreement; its terms are not evaluated.

Statement	Basis	Resolver	Does not show
Every dependency supplied by another legal entity has the supplier’s written information on file.	recomputed	dependencies.supplier_information_on_file	A pointer to the agreement, pinned by hash; its content is not evaluated.

Use in accordance with instructions

art26-1-use-per-instructions · Art 26(1) · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Deployers use the system in accordance with the instructions for use. Applies to: classification high; roles deployer, provider_and_deployer.

Statement	Basis	Resolver	Does not show
Every action outside the contract or grant was refused.	recomputed	authz.out_of_grant_refused	Covers actions the gateway observed; the grant itself is a declaration.
Sampled runs stay within the declared intended use.	judged	judge.per_instructions	Judged by a model over a disclosed sample; stated with human agreement, never merged with recomputed rows.

Human oversight assignment

art26-2-oversight-assignment · Art 26(2) · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Human oversight is assigned to natural persons with the necessary competence and authority. Applies to: classification high; roles deployer, provider_and_deployer.

Statement	Basis	Resolver	Does not show
A human can override or stop the system.	recomputed	oversight.override_available	Shows the controls exist and are assigned, not that the assigned people are competent.

Monitoring and suspension

art26-5-monitoring-and-suspend · Art 26(5) · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Deployers monitor operation and suspend use where the system presents a risk. Applies to: classification high; roles deployer, provider_and_deployer.

Statement	Basis	Resolver	Does not show
Liveness, drift and health were evaluated every day.	recomputed	monitoring.daily_health	—
Suspension and run-abort controls are available and every use of them was sealed.	recomputed	monitoring.suspend_available	—

Log retention (deployer)

art26-6-log-retention-deployer · Art 26(6) · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Deployers keep the automatically generated logs under their control for at least six months. Applies to: classification high; roles deployer, provider_and_deployer.

Statement	Basis	Resolver	Does not show
Logs are kept for at least six months (183 days).	recomputed	retention.floor {"floor_days": 183}	Record identifiers, statements and tree heads are permanent; bodies follow retention.

Information to workers

art26-7-worker-information · Art 26(7) · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Workers’ representatives and affected workers are informed before workplace use. Applies to: classification high; roles deployer, provider_and_deployer.

What the gateway cannot show: Organisational duty, evidenced by declaration only.

Statement	Basis	Resolver	Does not show
Worker information is declared.	declared	profile.field_declared {"path": "frameworks.eu-ai-act.worker_information_reference"}	Organisational duty; the information given is not evaluated.

Fundamental rights impact assessment

art27-fria · Art 27 · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Bodies governed by public law, private entities providing public services, and deployers of Annex III 5(b)/(c) systems assess the impact on fundamental rights. Applies to: classification high; roles deployer, provider_and_deployer.

Review: Applies to public bodies, public-service providers and Annex III 5(b)/(c) deployers — confirm whether the deployer is in scope.

Statement	Basis	Resolver	Does not show
A fundamental rights impact assessment is on file.	recomputed	evidence.pointer_on_file {"kind": "impact_assessment"}	A pointer to a document produced outside the platform; its content is not evaluated.

Registration in the EU database

art49-registration · Art 49, Art 71 · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

High-risk systems (and, for public bodies, their use) are registered in the EU database. Applies to: classification high.

Statement	Basis	Resolver	Does not show
The EU database registration identifier is recorded.	declared	docs.registration_recorded {"framework": "eu-ai-act"}	The identifier is declared; the registry entry itself is not queried.

AI-interaction notice

art50-1-ai-interaction-notice · Art 50(1) · applies from 2026-08-02

People interacting with an AI system are informed they are interacting with an AI, unless obvious from the circumstances. Applies to: profile flag core.interacts_with_natural_persons.

Statement	Basis	Resolver	Does not show
People were told they were dealing with an AI before the first substantive turn.	recomputed	notice.before_first_turn	Counts conversations through the Brutor portal (notice records); other clients must render their own notice.
The notice text and the surface it is shown on are declared and sealed.	declared	notice.text_declared	—
Sampled notices are clear and distinguishable.	judged	judge.notice_clear	Judged by a model over a disclosed sample; stated with human agreement, never merged with recomputed rows.

Marking of synthetic content

art50-2-synthetic-marking · Art 50(2) · applies from 2026-08-02

Synthetic audio, image, video or text outputs are marked in a machine-readable format. Applies to: profile flag core.generates_synthetic_content.

What the gateway cannot show: Content-level marking is not provided by the gateway.

Statement	Basis	Resolver	Does not show
The approach to marking synthetic content is declared.	declared	synthetic.marking_declared	Content-level marking (watermarking, C2PA) is not provided; outputs carry an X-Brutor-AI-Generated header only.

Synthetic-content marking for legacy systems

art50-legacy · Art 50(2) as amended by Reg. 2026/1744 · applies from 2026-12-02 · introduced in catalog 2026-09

Art 50(2) marking duties extend to systems placed on the market before 2 August 2026. Applies to: profile flag core.generates_synthetic_content.

Review: Applies to systems placed on the market before 2 August 2026 — confirm the placement date.

Statement	Basis	Resolver	Does not show
The approach to marking synthetic content is declared.	declared	synthetic.marking_declared	Content-level marking (watermarking, C2PA) is not provided; outputs carry an X-Brutor-AI-Generated header only.

Deep fake and public-interest text disclosure

art50-4-deepfake-text-disclosure · Art 50(4) · applies from 2026-08-02

Deployers disclose deep fakes, and AI-generated text published to inform the public on matters of public interest. Applies to: roles deployer, provider_and_deployer; profile flag core.generates_synthetic_content.

Statement	Basis	Resolver	Does not show
The disclosure approach for deep fakes and public-interest text is declared.	declared	profile.field_declared {"path": "frameworks.eu-ai-act.deepfake_disclosure_approach"}	Publication happens outside the gateway; the disclosure is not observed.

Post-market monitoring

art72-post-market-monitoring · Art 72 · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

A post-market monitoring system proportionate to the risks, fed by production evidence — the run ledger and drift findings are the raw material. Applies to: classification high; roles authorised_representative, distributor, importer, product_manufacturer, provider, provider_and_deployer.

Statement	Basis	Resolver	Does not show
Liveness, drift and health were evaluated every day.	recomputed	monitoring.daily_health	—

Reporting of serious incidents

art73-serious-incidents · Art 73 · applies from 2027-12-02 · earlier catalog: 2024-08 → 2026-08-02

Serious incidents are reported to the market surveillance authority: within 15 days of awareness; 10 days on a death; 2 days when widespread or critical infrastructure is affected. Applies to: classification high.

Statement	Basis	Resolver	Does not show
An incident register is kept and no open incident is past its deadline.	recomputed	incidents.register_exists	—
Serious incidents were reported within the Art 73 deadline (15 days; death 10; widespread 2).	recomputed	incidents.reported_within_deadline {"framework": "eu-ai-act"}	Deadlines are computed and shown, not enforced; awareness time is operator-entered.

High-risk duties for Annex I embedded systems

annex1-embedded · Annex I, Art 6(1) · applies from 2028-08-02 · earlier catalog: 2024-08 → 2027-08-02

High-risk duties where the AI system is a safety component of (or is itself) an Annex I regulated product — machinery, medical devices, vehicles… Applies to: classification high.

Review: Brutor cannot infer whether the system is embedded in an Annex I product — review and record the determination.

Statement	Basis	Resolver	Does not show
The Annex I / Annex III determination is declared.	declared	profile.field_declared {"path": "frameworks.eu-ai-act.annex"}	The determination is the operator’s; it is not evaluated.

GPAI model provider duties

art53-gpai · Art 53, Annex XI/XII · applies from 2025-08-02

Technical documentation, downstream information, copyright policy and training-content summary for general-purpose AI models. Attaches to the MODEL provider — this system is exposed through the GPAI models it binds. Applies to: systems binding a gpai model.

Statement	Basis	Resolver	Does not show
The GPAI provider’s downstream documentation is referenced.	declared	profile.field_declared {"path": "frameworks.eu-ai-act.gpai_documentation_reference"}	Model-provider duty; the documentation is not evaluated.

Systemic-risk GPAI duties

art55-gpai-systemic · Art 55 · applies from 2025-08-02

Model evaluation, adversarial testing, incident reporting and cybersecurity for GPAI models designated as systemic-risk. Attaches to the MODEL provider. Applies to: systems binding a gpai_systemic_risk model.

Statement	Basis	Resolver	Does not show
The GPAI provider’s downstream documentation is referenced.	declared	profile.field_declared {"path": "frameworks.eu-ai-act.gpai_documentation_reference"}	Model-provider duty; the documentation is not evaluated.

Related

• Framework registry & evidence bases — how catalogs, resolvers and the version in force work
• The Compliance console — the Obligations matrix for this framework
• Incidents, reports & documentation