The EU Artificial Intelligence Act (Regulation (EU) 2024/1689) requires high-risk AI systems to maintain automatic, lifetime logging of events sufficient to enable post-market monitoring and incident investigation. Article 12 of the Act, enforceable from 2 August 2026, specifies the obligation but not the technical mechanism. Three existing families of technical practice — content provenance standards, AI management systems standards, and trust-engineering frameworks — each address part of the obligation, but none in combination delivers the architectural properties required for compliant logging at the per-inference level for systems operating at scale. This paper identifies the specific architectural gap between Article 12 and the current state of practice, develops the Integrity Clash argument (Nemecek et al., 2026) into a general principle, and presents per-inference governance certification — a cryptographically signed certificate generated as a structural output of a non-bypassable execution-control layer — as a candidate technical specification that closes the gap. Four architectural properties are proposed as the minimum set required for Article 12-compliant logging: per-inference, architecturally bound, externally verifiable, and fail-closed. The candidate specification is described at a level of detail sufficient for standards-body consideration. UK patent applications relating to the mechanism are filed and on the public record; standards-essential claims are offered under fair, reasonable, and non-discriminatory (FRAND) licensing terms. This paper is written for BSI ART/1 working group members, ISO/IEC JTC 1/SC 42 contributors, EU AI Act Article 12 compliance leads, and technical leaders at regulated organisations preparing for the August 2026 enforcement deadline.
Christopher Hamilton (Tue,) studied this question.