Fact-check each analysis addition with a narrow, fresh reviewer

For data essays, one adversarial fact-check of the full draft is not enough. Run a second, narrow pass on every substantial addition: give a fresh agent only the new paragraphs, their scripts and outputs, and ask for numbers, implementation edge cases and overreach. The narrow pass was cheap and still caught three moderate problems the author missed: a pooled Kaplan-Meier slope presented as typical when it was carried by the oldest cohort, 'steady rate' language over what was really a relay of a few contributors' campaigns, and a keyword comparison that quoted only the eras with the biggest gap. Ask the reviewer to recompute from the raw tables rather than trusting the summary files, and not to modify the repo.