Researchers propose Publication Authority framework to make AI-assisted claims independently verifiable

Researchers propose a framework called Publication Authority to address ambiguities in AI-assisted claims, where evidence, analysis, human authorization, presentation, and correction history may refer to different states. The goal is to make these claims independently challengeable by enforcing exact-state, non-transferable, single-use publication capabilities.

The framework is instantiated in PAC-2026 (Publication-Accountability Calculus), a machine-readable candidate protocol under the AIJIM framework. PAC-2026 introduces six obligations that govern distinct aspects of the publication process: evidence, runs and artifacts, measurement disclosure, authorization, surface correspondence, and lifecycle continuity. Each obligation yields a target-bound witness, localized counterexample, or localized unverifiability; no single obligation can compensate for another.

The evaluation focuses on the fourth bounded semantic freeze (SF-4), a fixed-profile specification designed for replaceable bindings. In tests, ten models explored 110,764 safe reachable states, while 76 unsafe configurations produced the expected violation or observer countermodel. The protocol separates the evidence horizon from verification time, explicitly rejecting authentic but causally invalid authorizations.

The authors report that a reader surface passing its correspondence check cannot authorize publication unless the accepted record admits that surface. They also note that a historical predecessor path reproduced 17 frozen authorization-successor outcomes, and an in-house, instance-blind test matched all 183 scored expectations. Same-host package execution reproduced 240 archived observations. The results support internal coherence, bounded safety, fault sensitivity, and limited constructibility, but the authors caution that the protocol does not guarantee factual truth, general refinement, blind interoperability, field efficacy, or standards status.

The preprint is authored by Torsten Olivi Tiltack, Yifei Dong, Kun Yu, Xu Wang, Wei Liu, Jianlong Zhou, Ren Ping Liu, and Fang Chen, and is published as arXiv:2609.17631. It is not peer-reviewed and is presented as a preprint for community feedback.