Evaluating Evidence Sufficiency in Frontier Language Models

Artificial intelligence systems are often asked to answer questions from retrieved webpages, uploaded documents, research papers, and institutional records. In these instances, access to a relevant source is only part of the reliability problem. Particularly, the model must also determine what the source actually supports. That is, a claim may be directly supported, directly contradicted, or simply left unresolved by the available evidence. The third case is far easier to miss than the former two because an unsupported conclusion can be coherent, plausible, and even closely related to the source. This study tests whether two frontier language models can reliably distinguish among those three claim categories.