What problem does it solve? Academic manuscripts accumulate numbers that may drift from their underlying experimental evidence, leaving claims unsourced or mismatched before submission. This Skill mechanically traces every number in LaTeX sections and tables back to fingerprinted evidence files, assigning a verdict of matched, mismatched, or unsourced so no unsupported figure survives to submission. ## Core Features & Use Cases - Per-number trace and verdict: Extracts every numeric token from manus/secs/ and manus/tabs/, follows % src: comments and claim-ledger evidence links to fingerprinted mates/ entries, and assigns matched, mismatched, or unsourced verdicts. - Claim ledger management: Flips notes/claims.md statuses between drafted, verified, and unsourced based on audit results, and files one tasks/ item per failure routed to the owning skill. - Staleness detection: Runs import.sh --diff to detect drifted upstream evidence, tainting matches made against stale sources and blocking verification flips. - Use Case: Before a submission freeze, run the auditor over the full manuscript to confirm every reported result traces to fingerprinted evidence, receive a CLAIMS_<date>.md report, and get a task list of every number needing a fix. ## Quick Start Ask the AI to run stage-clms-auditor on the whole manuscript, or pass a section name or claim ID such as C7 to audit only that scope.
