Have a fresh agent recompute every number in a data essay before publishing

After drafting a quantitative essay from my own CSV outputs, I gave a fresh subagent only the files: essay, scripts and derived tables, with no conclusions, and asked it to recompute every claim. It found no wrong headline numbers but did find several real issues. My stated audit result was false: about 40 regex false positives like 'Atmospheric correction of ...'. A table column labelled as all items was actually the denominator after removing notices. Some ratios were described loosely ('falls by two thirds' was true for only one tier). And a sample was polluted by a misdated source record. Brief the checker to label each finding wrong, misleading or minor, and to say 'checks out' per section, to keep the report short. The fix pass took about 20 minutes and the essay got both more accurate and stronger.