Hallucination testing deliberately probes for confidently wrong output.
One method asks questions whose answers are absent from the context.
A correct system says it does not know rather than inventing something.
Another method plants contradictions and checks whether they are noticed.
Track hallucination rate as a first-class metric, not an afterthought.
Hallucination testing deliberately looks for confidently wrong output.
Waiting for users to find it is not a testing strategy.
The most useful probe asks questions the context cannot answer.
The right behaviour is to say the information is not available.
The wrong behaviour is a fluent plausible answer, and that is what you are measuring.
Include questions about things that do not exist at all.
A model asked about a fictitious policy will often describe it in detail.
Plant contradictions between two documents and see whether the model notices.
A good system flags the conflict rather than silently choosing one.
Ask for specifics such as dates, numbers and names, since those are where invention shows.
Check citations point at passages that genuinely support the claim.
Track the hallucination rate as a first-class metric with a target.
It should be reported alongside accuracy, not mentioned when someone complains.
Low temperature reduces it and does not remove it.
The most reliable reduction is giving the model permission to say it does not know.
One question on this card. Every option explains itself once you pick, right or wrong.
18 concepts in this module, then a graded test with every answer explained.