paper-claim-audit

To verify paper claims before submission, run a zero-context audit that traces every numeric and comparative statement to exact raw result files, outputting PASS/WARN/FAIL statuses and a PAPER_CLAIM_AUDIT.json artifact.

A zero-context audit reviews only paper sources and raw result/config artifacts without executor narratives or prior context, ensuring quantitative claim reconciliation remains free from confirmation bias.

Check for rounding drift and delta arithmetic errors by applying a rigorous mismatch taxonomy that detects exact mismatches, config mismatches, aggregation errors, and caption-to-figure inconsistencies during claim extraction.

Yes, you can audit scope overclaims and aggregation errors by running the audit on your current paper directory and associated raw results, which checks that every quantitative claim matches the corresponding raw result files.

The best way to generate machine-readable assurance outputs is to run an audit that always emits a paper/PAPER_CLAIM_AUDIT.json file plus a human-readable PAPER_CLAIM_AUDIT.md report for downstream automation.

You should not use it when executor narratives or prior audit context are the primary sources of truth, because the zero-context protocol strictly reviews only paper sources and raw result/config artifacts.