paper-claim-audit

Verifying paper claims against raw experimental data requires extracting every numeric assertion from .tex files and cross-referencing values directly against raw .json, .csv, .yaml, and metrics files to ensure exact matches without relying on executor summaries.

Zero-context evidence auditing prevents confirmation bias by running a fresh reviewer with no prior logs or narrative context, checking only paper .tex claims against raw evidence files to detect mismatches in numeric accuracy and rounding.

Detecting rounding errors or cherry-picked best-seed results involves tracing extracted quantitative claims back to raw result files, flagging rounding drift, aggregation errors, and best-seed versus average discrepancies for submission readiness.

Auditing LaTeX paper numbers against CSV and JSON files is supported by strict extraction of numeric and scope claims from .tex documents, cross-referencing them against raw .json, .csv, .yaml, metrics, and config inputs to confirm exact matches.

A paper claim audit outputs deterministic PAPER_CLAIM_AUDIT.md and PAPER_CLAIM_AUDIT.json files containing per-claim verdicts and hashes, providing structured downstream verification steps for submission assurance workflows.

Executor summaries should not be relied upon for submission readiness when confirmation bias risks exist, as quantitative claims require fresh-model zero-context review against raw evidence to catch caption mismatches, config discrepancies, and scope overclaims.