An AI evaluation infrastructure built on billions of synthetic persona records reports a high adherence rate — proof the model can play an assigned role. Article 10 of the AI Act asks whether the declared population is representative of real users, and who, independent of the producer, can confirm it. This article examines the gap between the two.

Certification of an AI system before deployment and proof that a specific decision was lawful at the moment it occurred are two distinct evidentiary claims, routinely conflated under the same vocabulary of compliance and audit trails. This article sets out where qualified timestamping, blockchain anchoring, and runtime governance stop, and where legal opposability begins.

The Digital Omnibus on AI defers the enforcement of high-risk obligations but leaves evidentiary exposure unchanged. When a system acts before the deferred deadlines, the governance state operative at T-0 remains the central question in civil, insurance, and contractual proceedings. The postponement affects regulatory temporality, not the requirement to hold an independently fixed, opposable governance record. The proof gap persists throughout the deferral window, closed only by a pre-execution cryptographic attestation architecture such as SOURCE 0.

Governance only becomes enforceable when three conditions are simultaneously met: T‑0 fixation, structural independence (S ∩ C = ∅), and legal opposability. Without these properties, no post-execution record can establish a governance state that was never sealed. This article articulates the architectural conditions under which governance ceases to be a management assertion and becomes a verifiable system property.