Hallucination & Fabrication Detectors

The adversarial counterpart to groundedness checking. Rather than asking whether claims are supported, it hunts specifically for invented entities, non-existent citations, and content generated without reference to the provided context at all.

High-consequence generation where a confident fabrication is worse than a refusal — legal, medical, financial and regulatory contexts.

Where the cost of a false positive exceeds the cost of a fabrication. Over-aggressive detection produces refusals on correct answers, which trains users to bypass the system.

Detected fabrications must block rather than warn where the output enters a record. A warning a user can dismiss is not a control.

Plausible fabrications evade detection precisely because they are plausible. Detectors reliably catch absurd invention and reliably miss the invented figure that sits within the expected range.

An audit response drafted from evidence records, where the detector catches a cited document reference that resolves to nothing in the evidence store.