From Record to Finding: Why Tamper-Proof Logs Cannot Establish Legal Oversight of Agentic AI

A supervisor charged with overseeing an agentic AI system must, under EU law, be able to establish certain findings of fact: whether personal data of a protected class left a controlled environment, whether a human had the authority and the window to intervene, whether an information barrier held, whether a delegated authority was valid at the moment it was used. This paper asks a narrow and prior question. Under what conditions can a runtime record answer such a finding at all? It develops an evidentiary-adequacy criterion: a representation can answer a determination of this kind only if it carries two things, the legal typing that maps recorded events to the operative category, and the relation, usually provenance or authority, on which the determination's truth depends. A representation carrying neither can raise a suspicion but cannot establish a finding. The criterion is a necessity claim, not a sufficiency claim; sufficiency fails for reasons the paper makes explicit. The criterion is stated formally, its necessity is shown by construction across a defined class of determinations, the property set is shown to be minimal, and its adequacy is argued to be bounded to that class rather than universal. The determinations in scope are characterised by a stated selection rule: they are binary findings of fact about specific events and their relations, not evaluative or statistical properties such as fairness or robustness. The criterion is supported, not derived, by three independently established results that converge on the same necessity from different directions: Ashby's Law of Requisite Variety, the Good Regulator Theorem, and the trace-versus-hyperproperty boundary of runtime verification. The paper is careful to claim no novelty in that convergence; the inversion of cybernetic model-dependence onto the overseer has been made before, notably by Aguirre (2025), as an impossibility argument about control. The contribution is the criterion itself, its instantiation in EU AI Act oversight, a gap analysis showing that existing governance frameworks do not answer these determinations, and a pre-registered experiment that tests the criterion against expert judgement. The legal claim is stated conditionally. A record that is not semantically and relationally interpretable cannot serve as evidence that effective oversight under Article 14 was possible, even where it satisfies the Article 12 duty to log. The structure recurs in any domain where legal findings depend on the execution paths of autonomous systems, of which EU AI Act oversight is the worked instance.