Independent verification protocol proposed for AI-assisted claims

Summary

The article proposes 'Publication Authority,' a formal, single-use, and non-transferable credential for AI-generated claims. This authority is concretized as the PAC-2026 protocol. The protocol includes six irrevocable obligations: evidence, execution, measurement disclosure, authorization, surface alignment, and lifecycle continuity. These obligations were verified through finite state model auditing over 110,764 accessible states. The work is currently a preprint specification and is not a deployed system or a peer-reviewed standard.

Why it matters

As AI-generated content floods publications and reports, the gap between what a model produces, who authorized it, and what was actually published creates a genuine accountability issue. A machine-readable and falsifiable publication record could allow third parties to challenge the origin of specific claims. This enables debate to proceed on concrete claims rather than abstract outputs. However, since the protocol is still an unadopted candidate specification, its practical impact is currently zero.

Highlights

• The six obligations are strictly irrevocable. Passing five obligations cannot compensate for a failure in the sixth. Only a record where all obligations are completed authorizes a single atomic publication transition.
• Verification is entirely endogenous: 10 finite models, 110,764 safe states, 76 expected unsafe configurations, 183 internally scored expectations, and 240 archived observations; all were self-generated.
• The authors explicitly disclaim factual accuracy, general improvement, blind interaction capability, field efficacy, and standard status. They position the work solely as a limited security and consistency audit.

Critical take

Model auditing over 110,000 states is necessary but absolutely insufficient validation for a protocol designed to manage real-world AI publications. The authors themselves admit a lack of field efficacy and interaction capability. All evaluation was conducted internally, with no independent replication. Consequently, the claim of 'falsifiability' remains somewhat circular at this stage.