ClaimScope
ClaimScope: Label-Blind Witness Verification for Mathematical Claim Auditing
A 72-task benchmark for auditing mathematical claims via verdict, scope, counterexample, and repair across optimizer, recurrence, and linear algebra domains.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info claimscope
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
1 sourced entryEvery number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
ClaimScope comprises 72 tasks covering optimizer-state transport, affine recurrences, and linear algebra. It requires models to provide a verdict, scope, counterexample (witness), and repair for each claim.
Scoring evaluates safe completion (valid structure, correct verdict/scope/repair, verified witness) and strict completion. Verifier feedback improved safe completions to 25/72 compared to 20/72 for generic feedback.