Every content department needs a gate between 'written' and 'published', and the gate has one question: which sentences state facts? AI drafts weave invented statistics and confident specifics into perfectly good prose — the 73% that came from nowhere, the study that doesn't exist.
The discipline: before publishing anything AI-drafted, extract every factual claim into a list and mark each one verified, unverifiable, or decorative. Cut the unverifiable, source the verified, and be honest that 'studies show' with no study is fiction.
The seed asks two models to audit the same paragraph for claims. Compare the audits: which model caught more? Which invented a verdict? QC work is itself a task you can race models on.