Invariant Relation and Non-Substitution: Authorization requirements in AI-assisted workflows

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

An artificial agent may correctly predict a person's agreement while lacking authorization to act. This paper develops a conditional non-substitution principle: where a justified task requires an actual event or relation, a prediction alone does not satisfy that requirement. It distinguishes types of relation and gives an elementary result about certification from readouts that discard a required event. A worked report-publishing example follows authorization through prediction, planning, and execution. It separates technical capability from a current mandate, preserves legitimate delegation, and makes the authority and timing assumptions explicit. An executable supplement enumerates 192 constructed attribute assignments and ten temporal examples. A conventional authorization baseline and the typed checks make identical decisions throughout; deliberately defective policies expose omitted distinctions. These are conformance checks on a stipulated model, not language-model experiments or evidence of superiority. Access control, usage control, assistance games, and normative-interface accounts already supply relevant resources. The contribution is a concrete specification-review procedure connecting an event-sensitive end to the representations used to authorize action. Its normative premises, practical adoption value, authentication assumptions, and implementation beyond the toy example remain separate burdens. Research status: Conceptual working paper; not peer reviewed. The supplement contains deterministic checks of a fictional authorization model. No language-model or participant experiment is reported. Version and contents: Paper 1 of the Invariant Relation working-paper series, version 1.1. This version revises an unpublished version 1.0 draft. The complete package contains the paper in PDF, HTML, and Markdown, a README, executable Python supplement, generated results, license notices, and a file manifest. Licenses: The paper and research materials use CC BY 4.0. The Python demonstration code and its software documentation use MIT. These licenses apply to their specified components, not as interchangeable options for every file. See LICENSE.md and supplement/LICENSE.txt in the complete package.