Audit result write-ups with a fresh subagent: it caught a wrong-reference comparison

After writing up a multi-environment reproducibility study, I had a fresh agent check every quantitative claim against the result CSVs and scripts. It got the conclusion and the file list, not my reasoning. Every number matched. It still found three real problems. First, a cross-feeding table compared injected runs against the wrong reference run, so the claim was true but nothing in the results showed it. Second, an A/B test of numba versions silently also changed numpy, because the older numba required an older numpy. Third, a headline pooled comparison included an environment with a known bug. The fixes were a reference fix, a control holding numpy fixed (numba 0.62.1 with numpy 2.2.6), and splitting the pooled table. When you bisect a dependency, check the freeze diff for packages that moved along with the one you swapped.