Tool · runs in your browser, nothing is sent anywhere · decision-support only, not legal advice

Live triage

The evaluated engine (Cycle 2: synonym normalisation plus the persons-net), ported line for line and drift-tested against the Python artifact on every benchmark row and a generated probe corpus. Describe an AI system in a sentence or two and get the complete decision record the paper calls a compliance passport. The rule set is English-lexical: the four-language corpus measures that limit, it does not remove it.

Reading the record

Candidate outcomes are lexical matches, never legal findings: a prohibited- or high-risk-candidate is always gated for confirmation. INDETERMINATE is a workflow state routed for review with the candidate provisions and the facts a reviewer would need; it is not a risk tier. Minimal means no Art. 5 / Art. 6(2) / Art. 50 trigger was detected within this triage — other routes, Annex I product safety above all, were not assessed.

The counterfactual is computed under the engine's own rule-state intervention — remove the decisive matched rules (and the persons-net where it is co-decisive) and re-apply the priority logic — so it is faithful to the decision procedure rather than a guess. Two things stay orthogonal by design: Article 50 flags survive any tier change (Art. 50(6)), and GPAI status is a model-level regime (Chapter V), reported alongside the tier, never as one.

Rule set v0.3.0 pinned to Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744. Reference labels in the evaluation are author-assigned. This is a baseline; on the post-freeze set it silently under-classifies 40% of systems.