From Warranted Acts to Revisable Authority: Practical Judgment, Critique, and the Continuing Justification of AI Delegation — A Critical Synthesis within the Answer-Entitlement Program

Abstract

This paper is the third in the author's answer-entitlement series. The first asked when an AI system is warranted in answering; the second asked what infrastructure supports bounded, recoverable delegation. This paper asks whether a sequence of individually warranted and recoverable acts can justify the continuing arrangement that selects their purposes and the criteria by which they are judged, and answers that it cannot. In doing so it withdraws three sufficiency readings of the program's own earlier position: reversibility as a permission rule, convergence or coherence as justification, and the aggregation of warranted acts into a justified arrangement. An AI system can give warranted answers and perform authorized, recoverable operations while the arrangement in which it works remains unjustified. The paper argues that where a consequential arrangement defended as supporting a person's agency helps shape the endorsements, preferences or evidence later cited for its continuing authority, those items cannot by themselves certify that authority unless their admissibility remains open to the person's challenge, directly or through a representative, by an effective route whose outcome does not depend solely on the criteria in dispute. The constraint applies a structure that Forst states for power in general, and Abukar for classification evidence, to endorsements an assistant helped produce; Lazar places AI's shaping of beliefs and desires within governing power. Building on work holding that legitimate acts do not aggregate into legitimate continuing authority and that power must answer to those subject to it, the paper offers a critical synthesis and internal correction of the program, not a new general theory. Of the three readings it withdraws, one was stated as a rule in the program's position paper and two were invited by its earlier formulations. It states five procedural specifications as design hypotheses: handoffs that reset neither a challenge nor its elapsed time; a terminal record of failed review that cannot be relabelled as a merits decision; a missed containment-review deadline treated as failure, not permission; continuity triage protecting separable assistance; and longitudinal comparison of the arrangement against a reference not controlled solely by the disputed profile. Constructed cases test the distinctions. The paper reports no empirical measurement of its own; §10.5 summarizes the confirmatory technical results of a preregistered companion study now in preparation, which, on its deposit, will carry the protocol, code and data: that study has confirmed the specifications' implementability and specified behaviour under scripted trajectories, and their normative work in real institutions remains untested. **Keywords:** AI delegation; answer entitlement; revisable authority; contestability; right to justification; influenceable preferences; endorsement; practical judgment; critique; human-AI interaction; MOBIUS. Release type: conceptual paper, a critical synthesis within the answer-entitlement program; the paper reports no empirical measurement of its own, and §10.5 summarizes the confirmatory technical results of the preregistered companion study (Toeda 2026g), which, on its deposit, will carry the protocol, code and data. AI disclosure: versions 0.1 to 0.4 were drafted with an OpenAI Codex agent (model gpt-6-astra, per the session record), which also used a locally run Gemma 4 26B-A4B model to locate passages in source documents, and version 0.5, including its argument, was revised with Claude Opus 5.5 (Anthropic), both under the author's direction; after the fifth review round the subtitle and the abstract were revised with Claude Fable 5.1 (Anthropic), also under the author's direction. Revisions in versions 0.2–0.4 responded to review comments that the author obtained from ChatGPT (OpenAI) and supplied to the drafting agent. The prior-art search, three quality-assurance rounds and the adversarial review of version 0.5 (fourteen passes in eight rounds, each run by a separate agent with its own context; each of the first three rounds had three passes, on the philosophical argument, on sources and quotations, and on scope and consistency, and the fourth, fifth, sixth, seventh and eighth had one confirmation pass each on the changes made after the round before; the third round reported no blocking finding, the fourth found an invalid step in a reply to an objection added after the third, the fifth, run on the corrected reply and on this statement, reported no finding, blocking or otherwise, the sixth, run by a Claude Fable 5.1 agent on the revised subtitle and abstract and on this statement, reported no blocking finding, the seventh, run by a Claude Fable 5.1 agent on this statement after the source of the review comments had been named, reported no finding, blocking or otherwise, and the eighth, run by a Claude Fable 5.1 agent on this statement after it had been moved to its own section with a summary in the header, and on the typesetting of the deposit files, reported no finding on the text and one typesetting defect, a markup label left unconverted in the deposit files, which was corrected before deposit) were carried out by other Claude agents, Claude Opus 5.5 except where stated, which are not independent in the way human referees would be; no human referee has reviewed the text. For version 1.0, the confirmatory results of the companion study (Toeda 2026g) were integrated into §10.5, and the related wording changes were made, by a Claude Fable 5.1 agent under the author's direction. Adversarial review of version 1.0: the version 1.0 changes went through five further passes in three rounds, each run by a separate Claude Fable 5.1 agent with its own context; the first round had three passes, on the numbers and protocol fidelity of §10.5, on scope and consistency, and on the whole text as an editor would read it, and found five blocking findings (correct counts attached to the wrong unit for two mechanisms, one over-general confirmation clause, the header's review count, and the wording about the companion's deposit) and twenty-nine minor ones, all of which were applied; the second round was one confirmation pass on those changes, which found one blocking finding, the sentence on the external timestamp of the companion's protocol, which placed the Bitcoin attestation before the confirmatory run although the block was mined after it, and four minor ones, all of which were applied before deposit; the third round was one confirmation pass on those corrections, on this statement and on the typesetting of the deposit files, which reported no finding, blocking or otherwise. The author takes responsibility for the deposited text. Working method only; the registered author is the human author alone. The record contains the paper as a PDF. Empirical companion (preregistered; in preparation): 10.5281/zenodo.23110635.