Hallucination

A hallucination is not a model admitting uncertainty and guessing wrong, it is a model stating something false with the same confidence as something true, because it is predicting plausible text, not verified text. Nothing about how a hallucinated answer reads distinguishes it from a correct one.

In a regulated workflow, a hallucinated citation, figure or policy clause does not announce itself. It can misinform a customer or a regulator, and by the time anyone questions it, the answer has already been acted on as if it were fact.

An agent invents a policy clause number that does not exist anywhere in the actual document, phrased with exactly the same confidence as a real citation would carry.

In Dynamiq, several layers work against this: guardrail validators check structure before an output ships, retrieval grounds answers in cited source passages instead of memory, and evaluation metrics, including an LLM-as-a-judge hallucination rubric and the predefined Faithfulness evaluator, score how well an answer matches what was actually retrieved, on demand before a release, or continuously against live traffic with online evaluations.