Fact-checking a paper against code: check what a test actually compares, not its count

When auditing a draft paper or README against the repo it describes, the most common overstatements were not wrong numbers (recompute all arithmetic; it usually checks out) but mis-scoped ones: (1) a test cited as "N windows compared against brute force" actually ran a different query path and a collapsed (last-value-wins) comparison, so it did not test the fix the sentence was about; open the test file and read the loop. (2) "corrupt files are rejected": a bit-flip test in the same repo showed ~1/3 of in-frame flips decode silently when there is no checksum; structural validation is not integrity. (3) a validation-cost figure quoted for one path (block decode +6-10%) while a sibling path cost much more (+29-35%). (4) a component the spec labels proposed being described as implemented. (5) a results table whose rows came from two different code variants (fast encoder on one host, reference encoder on another). (6) geomean over per-item ratios reported without the aggregate ratio, when one outlier dominates. Workflow: grep the spec for Implemented/Proposed labels, read tests/ for what each assertion compares, recompute every ratio and total with a one-line python, then tabulate claim -> OK/WRONG/UNSUPPORTED/OVERSTATED with the source line.