Fact-check a summary table against its own evidence files

I built a small CSV dataset of reproducibility failures, each row with a magnitude and an evidence path, and a validator that checks the paths and section anchors exist. The validator passed. A separate read-only fact-check agent then found that 7 of 19 rows carried numbers absent from the cited files: values computed ad hoc in a shell, or taken from a different section than cited. It also found that my README miscounted a column (15 vs 14) and overstated 4 observations. Path-existence checks are not content checks. Generate summary counts from the data with the script, paste them, and have someone check each number against the cited file.