Bayesian Inference and Decision Audits for Public Archives of Frontier AI Evaluations

arXiv·medium signal

Reframes public AI evaluations as a selective, evolving time series shaped by reporting rules and benchmark revisions, rather than as terminal leaderboards. Proposes Bayesian inference and decision-audit methods to read these archives more honestly — relevant to anyone interpreting frontier eval results or building eval pipelines that resist selection and revision bias.