The Authorization Threshold: When an AI Action Qualifies as Pre-Execution Authorized

This paper defines the pre-execution authorization threshold for AI systems. "AI governance" names many functions: monitoring, model alignment, content filtering, audit logging, policy administration, and runtime enforcement. These functions are not interchangeable, and the paper does not rank them. It isolates one narrower, action-specific question: when can a specific covered effect-bearing action be classified as pre-execution authorized? The controlling test is composite, an interpretive restatement of the Authorization Artifact Test. Before the covered effect occurs, a runtime authorization boundary must emit an action-bound verdict represented in an authorization artifact; and an independent third party must be able to reconstruct that verdict from the artifact and its authenticated bound materials under a declared replay mode, without access to the governed system. Passing this threshold establishes neither complete governance nor legal compliance: Output Integrity, Input Integrity, Replay Integrity, non-bypassability across the declared topology, and 5TS conformance each require separate evidentiary support, and every conclusion remains bounded by the declared scope. Along the way, the paper classifies components by demonstrated function rather than product label, applying the Composition Test to observability systems, models, policy engines, gateways, and human resolvers; separates observability evidence from authorization evidence under the three permissible evidentiary conclusions (the evidence supports the claim, a failure witness defeats it, or the claim is not established); treats constraint representation as an authorization input obligation rather than a separate verdict function; states the runtime authorization boundary semantics for ALLOW, DENY, and ABSTAIN, with authorized resolution operating through the boundary and emitting a separate resulting action-bound verdict for the held action; replaces the earlier standalone validity and completeness inventories with a doctrine crosswalk onto the controlling instruments; specifies the authorization artifact together with its authenticated bound materials; and closes with a vendor-neutral evaluation sequence. Provenance is treated with precision: it establishes the origin of verdict-grounding inputs, origin rather than truth, and is necessary but not sufficient for an Input Integrity claim. This work is intended for researchers, regulators, enterprise buyers, compliance leaders, and technical architects who need to evaluate whether a specific action was subject to pre-execution authorization rather than only retrospective observation or review. Part of the FERZ deterministic governance research corpus. Within that corpus, this paper is the interpretive synthesis: it introduces no new conformance standard and no new test beside the Authorization Artifact Test, the Authorization Boundary Integrity Model, and the Five Tests Standard. It maps the threshold onto those instruments together with the ABIM Evidence Requirements, Execution-Time Authorization, the Closed-World Bargain, and the Composition Test, and identifies the point at which their requirements support the bounded conclusion that a specific action crossed the pre-execution authorization threshold. Update (August 2026, v2.0): Version 2.0 narrows the authorization threshold from a general classification of governance to the specific question whether a covered effect-bearing action qualifies as pre-execution authorized. It replaces the Version 1.1.0 Class A/Class B taxonomy, evidence-versus-proof dichotomy, seven validity conditions, and five completeness dimensions with the controlling AAT, ABIM, 5TS, ABIM Evidence Requirements, runtime authorization-boundary, closed-world, and composition framework; distinguishes an authorization artifact from its authenticated bound materials; separates AAT classification from Output, Input, and Replay Integrity; corrects ABSTAIN and authorized-resolution semantics; conditions determinism and reconstruction on declared decision state and replay mode; and removes claims that successful verification establishes correctness or legal compliance. Version 2.0 supersedes Version 1.1.0 (July 2026; edition DOI 10.5281/zenodo.19270987) but preserves the original authorization-threshold intuition as an interpretive synthesis rather than a separate conformance standard.