Verifiable method · Versioned policy

Method, scope and sources

How the portal turns answers into an outcome, which evidence it preserves and where operational support ends.

Deterministic rules

Primary outcomes derive from explicit, versioned conditions. An LLM does not decide compliance.

Evidence by design

Answers, policy, disclosure, hashes and history form an addressable and verifiable snapshot.

Human review

Ambiguity, exceptions and sensitive contexts are stopped and assigned to a competent reviewer.

Logic covered by the engine

The classifier covers scope, roles, exclusions, prohibited practices, high-risk, GPAI, Article 50, FRIA and cross-cutting obligations.

Scope, roles and high risk

It qualifies the system and operators, then checks Annex I and III paths without replacing competent assessment.

Prohibited practices

Potential Article 5 prohibitions stop positive handoffs and require legal escalation while keeping all other findings visible.

GPAI and transparency

It distinguishes GPAI-provider duties and Article 50 cases; the specialist module remains separate and bound to the exact revision.

Exceptions and FRIA

Exceptions, draft guidance and FRIA applicability are recorded for manual review, not converted into automatic approval.

Primary sources

References linked to the active release of policy eu-ai-act-classifier.

Scope and limitations

The portal organizes an assessment and documentation process but does not automatically guarantee compliance. It does not fully cover copyright, privacy, advertising, digital services, electoral law or sectoral rules. EU icons are voluntary; the disclosure obligation, where applicable, derives from the Regulation. High-impact or disputed cases require a qualified professional.

eu-ai-act-classifier@2026-08-04.1 · eu-ai-act-classifier-guidance@2026-08-04.2