The Named Set Error: A Specification for Correcting Prohibition Rules Given to AI Agents

This specification defines a failure mode in prohibition rules given to AI agents — the Named Set Error, in which a rule meant to exclude a whole class of actions names some of its members instead, inviting the reading that everything unlisted is permitted — and a two-sentence repair. It states the repair in normative form (a preservation clause restoring all standing prohibitions, and a routing instruction for residual uncertainty), partitions the responsibility the repair does and does not assign, and provides an afternoon-scale validation protocol with five registered, falsifiable predictions, including an independent subtest for the proposed salience ("lamp") effect. The reference incident is the July 18, 2026 fabricated tip submitted by an AI model to the Philadelphia Police Department's public tip form, per Anthropic's October 9, 2026 report and the department's public statement of the same date. The companion essay, "Name, Set, Match," presents the error and repair in narrative form.