01 · Outcome
What decision will the answer support?
Define the user, task and consequence rather than beginning with a general chatbot brief.
Technical evaluation
Bring an answer you must defend.
A useful evaluation starts with a real decision, an approved source set and a failure condition. We demonstrate what the system answers, what it refuses and what evidence remains afterwards.
Evaluation path
Five questions before a pilot.
02 · Corpus
Which sources are approved?
Identify ownership, versions, access boundaries and the conditions under which a source becomes current.
03 · Refusal
When must the system stay silent?
Agree the unsupported, contradictory and out-of-scope cases before measuring answer quality.
04 · Evidence
What must be inspectable later?
Specify the claim, citation, version, verdict and record needed for review.
05 · Success
How will we know it worked?
Use a test set and explicit thresholds for support, refusal, citation fidelity and reproducibility.
Next step