Researchers presented the PAC-2026 protocol for verifying AI publications

A group of researchers published a paper describing PAC-2026 (Publication-Accountability Calculus) — a protocol designed to ensure the verifiability of statements made using AI. The system introduces the concept of "publication authority" as an atomic, non-transferable action that requires a complete chain of evidence to confirm the authenticity of data.

The protocol is based on the SF-4 (bounded semantic freeze) specification, which imposes six mandatory conditions on the publication process. They cover the evidence base, artifact history, disclosure of measurement methods, and continuity of the model's life cycle. The system excludes the possibility of compensating for one missing parameter with another: a complete and correct set of records is required to complete a publication transaction.

During the testing of the protocol, researchers analyzed 110,764 secure system states. In 76 cases, configurations were deemed insecure, leading to expected disruptions or generation of countermodels. Internal tests showed that the protocol successfully reproduces 17 frozen authorization states and 240 archival observations, ensuring strict consistency between the surface representation of data and their actual history.

Developers emphasize that PAC-2026 does not guarantee the actual truth of the content, but ensures its verifiability and accountability. The protocol separates the horizon of evidence from the time of verification, preventing the publication of data that has passed authentication but lacks a causally valid authorization history. This solution is aimed at combating the problem where AI statements appear authoritative but rely on disparate and unrelated data states.