The Override Asymmetry: Why ABSTAIN-Plus-Human-Override Is Not Guardrails-Plus-Human-in-the-Loop

Abstract

Review-only human-in-the-loop and ABSTAIN with authorized resolution look alike in a workflow diagram. Both place a human near the moment an AI system might act, and both interrupt the action's path to execution. This technical note shows that the similarity is an artifact of diagram resolution and states the rule that separates the two. A human decision becomes part of authorization only when a non-bypassable, fail-closed runtime authorization boundary materially consumes the human's authority-bound input to emit the action-bound verdict on which execution depends, and that verdict is represented in an independently reconstructable authorization artifact. This is the note's operational reading of the Authorization Non-Substitution Principle for the case of human participation. A human supplies input; the boundary emits the verdict; the authorization artifact represents the verdict. Participation that does not meet the rule may support an authorization process, but it does not itself constitute authorization. The two arrangements differ on three axes: the basis on which the human is invoked, how the human's determination enters the architecture, and the properties of the record it leaves. The second axis carries the argument and separates into two consumption events. First-order consumption, the boundary's consumption of the human's input, is required for the held action to be authorized at all. Second-order consumption is optional: where a separately authorized policy-amending pathway exists, it may consume the resulting authorization artifact to change the policy governing later actions in a covered class. That second act changes future policy. It is not what authorized the held action. The note states workload functions over fixed-policy intervals for both arrangements, makes no claim that either is cheaper, and restricts its only rate claim to a covered recurring action class under declared conditions. Reserved action classes, in which every action returns to a human by design, fit the model without qualification. The note describes how the two arrangements degrade under sustained volume, distinguishing an arrangement that degrades without authorization-grade evidence from one that degrades traceably with respect to policy, authority, and verdict, and it treats drift as a separate analytical judgment over that record rather than something the artifacts render. The note is situated within the FERZ research corpus and provides the operational reading of Criterion 6 (Override Governance) from On the Impossibility of Observability-Based Authorization: what it means for an override to occur within the enforcement architecture rather than around it. It draws on the Authorization Artifact Test, the Authorization Boundary Integrity Model, the Five Tests Standard, Authority versus Authorization, and The Closed-World Bargain, and includes a doctrinal mapping to the five tests and the Composition Test. The mapping is not a conformance claim about any implementation. The argument is structural and uses a stipulated comparator. It makes no empirical claim about deployment economics or policy maturation, and no claim about any particular implementation, including any FERZ implementation. Version history. v2.0 (August 2026): supersedes the April 2026 formulation. Treats codification into policy as optional rather than definitional; separates action-scoped authorized resolution from separately authorized policy amendment; states that an unresolved ABSTAIN remains ABSTAIN and is not converted to DENY; restates the workload comparison as interval functions with a restricted rate claim; enforces the input, verdict, artifact sequence throughout; aligns terminology to the authorization-artifact class term and independent reconstructability; updates all corpus references. Version 1.4 remains on this record as the April formulation. v1.4 (April 2026): added Appendix A, Boundary Conditions. Citation and license. Copyright © 2026 FERZ, Inc. This work is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). Cite the concept DOI 10.5281/zenodo.19772248 for the work and the version DOI 10.5281/zenodo.22037126 for this edition. Canonical page: ferz.ai. Full corpus: the FERZ community on Zenodo.