Answers are grounded in this paper and cite their sources.
A pre-execution check can be real and still not establish authorization. The distinction lies in what the check evaluates, what governing information it possesses, whether execution depends on its verdict, and what an independent reviewer can reconstruct afterward. Checked Is Not Authorized examines the promotion of operation-level access controls into broader claims of action authorization for autonomous AI systems. Credential brokering, per-command policy evaluation, privileged access management, and session recording can provide legitimate security controls. Their presence does not, by itself, establish that a consequential AI action was authorized under the governing conditions applicable to its effect. The article develops three requirements for substantiating a pre-execution action-authorization claim: Governing-information sufficiency: The evaluation must address the policy, authority, purpose, and relevant conditions material to permission for the proposed action. Authorization-bound execution: Release must depend on a valid, action-bound ALLOW through a non-bypassable, fail-closed runtime authorization boundary within the declared scope. Independent reconstruction: The retained authorization artifact and its bound materials must support reconstruction of the historical verdict without reliance on the governed system or the vendor's live decision service. A worked example involving an AI maintenance agent and a database-backup preservation hold illustrates the conditional information-sufficiency result: where an evaluator receives identical information for situations requiring different permission outcomes, it cannot guarantee correct resolution in both. Deterministic evaluation and pre-execution timing do not supply missing governing facts. The article then examines complete mediation, enforcement independence, fail-closed recovery, input provenance, and the distinction between policy-evaluation records and independently reconstructable authorization evidence. Its practical contribution is a buyer-oriented acceptance sequence. It applies the six published criteria of the FERZ Enforcement Test Protocol to a declared deployment scope, followed by separate demonstrations addressing selected material governing conditions, independent reconstruction, and input provenance. These additional demonstrations do not replace or amend the normative protocol. Successful tests establish only what was demonstrated within the tested configuration and scope. Audience: Enterprise security architects, Chief AI Officers, risk and compliance executives, procurement officials, auditors, regulators, and technical diligence teams evaluating AI-agent authorization claims. Scope: This publication is technical commentary and applied analysis grounded in previously published FERZ research. It does not introduce a new formal theorem, report empirical deployment results, or evaluate a named vendor. Its criteria apply to FERZ's own authorization claims as well as those of other suppliers. Research foundations FERZ Technical Advisory TA-2026-01: The Authorization Artifact Gap On the Impossibility of Observability-Based Authorization Observability Is Not Enforcement Five Tests Standard (5TS) Enforcement Integrity in Pre-Execution Authorization Canonical article: https://ferz.ai/articles/checked-is-not-authorized
No takes yet. Share an insight, caveat, or question.
Edward Meyman (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: