ARC method boosts changepoint detection in AI models

A new arXiv paper introduces ARC (Augmented-Rank Conformalization), a family of rank-based scores for changepoint localization that guarantees finite-sample coverage for every frozen weight configuration, including random initialization and mistraining. The authors prove that the entire ARC confidence set is almost surely invariant under strictly increasing monotone transforms, so certified set lengths hold across re-expressions—unlike plug-in scores, whose lengths change. Simulations confirm nominal coverage for all ARC scores, including sabotaged networks, identical sets under monotone transforms where plug-in scores inflate, and on the well-log benchmark ARC localizes annotated shifts to three to five candidates and flags misfit with an empty set.