info snack
Hallucination
Confident output that is not grounded in the task, sources, tools, or real world.
companion card
A hallucination is confident model output that is not grounded in the task, sources, tools, or real world.
What it means
A hallucination can be a fake fact, a made-up citation, an invented tool result, or a plausible answer that the available evidence does not support.
Why product teams care
The fix is rarely just changing temperature. Better grounding, retrieval, schemas, abstention behavior, citations, and evals usually matter more.
Understudy angle
Understudy can track hallucination-prone examples and score candidate fixes against a groundedness rubric.
take this with you
Hallucination is a system problem: context, tools, policy, evaluation, and UI all matter.
Add an explicit 'not enough evidence' pass condition to one eval.