Most evaluations of mathematical agents check whether a final answer is correct or a proof is accepted. ClaimScope takes a different route: it asks an agent to audit a proposed claim by giving a verdict, clarifying its scope, providing a counterexample if needed, and suggesting a repair. The new benchmark comprises 72 tasks built around that protocol.
The abstract positions this as a complement to existing metrics rather than a replacement. Because the source abstract is truncated after "Tasks cover…", the specific areas covered by the 72 tasks are not described in this material. Even so, the core idea is clear: mathematical reasoning should be judged not just by producing a correct result, but by critically examining claims.