paper-claim-audit

Paper-to-evidence verification checks every quantitative claim against raw result files using a zero-context reviewer to reduce confirmation bias. It applies strict matching with standard rounding rules and validates relative-delta arithmetic to ensure paper-to-results fidelity.

Zero-context verification means a fresh reviewer with no prior context checks quantitative paper claims against raw result files. This method prevents confirmation bias by excluding executor summaries and relying only on paper sources and raw artifacts.

You audit paper tables and captions by reconciling each number and scope statement against raw results and config artifacts. The audit emits structured per-claim verdicts that identify exact matches, rounding-allowed matches, ambiguous mappings, and missing evidence.

Yes, claim auditing works for specific datasets and configurations while excluding executor summaries. The process reads only paper .tex sources and raw results or config artifacts to verify paper-to-results fidelity without being influenced by executor summaries.

A paper claim audit identifies exact matches, rounding-allowed matches, ambiguous mappings, missing evidence, config or aggregation mismatches, number mismatches, and unsupported scope overclaims. These are emitted as structured per-claim verdicts in both human-readable and machine-readable formats.

Auditing paper numbers generates a human-readable markdown report and a machine-readable JSON file. These deterministic audit artifacts contain structured per-claim verdicts with PASS, WARN, or FAIL outcomes for downstream verification workflows.