Making AI-Assisted Claims Independently Challengeable: Publication Authority and a Protocol for Falsifiable Publication Records

The paper introduces Publication Authority, a single-use, non-transferable capability that ensures AI-assisted claims can be independently challenged by providing a machine-readable, falsifiable publication record. It presents the PAC-2026 protocol, evaluates its fourth bounded semantic freeze (SF-4), and demonstrates through extensive modeling that the system enforces strict obligations on evidence, authorization, and lifecycle continuity. The study confirms internal coherence, bounded safety, and fault sensitivity, though it does not address factual truth or field efficacy.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.

• safety