Choose representative examples
Use examples that reflect real work without exposing information you cannot share. Include routine tasks, unusual inputs, missing context, and cases with no valid answer. Record why examples were chosen; a convenient sample can miss important failures.
Define review criteria
Specify what a useful answer contains and what makes an output unacceptable. Consider correctness, completeness, disclosure risk, and verification time. When a workflow can take actions, assess authorisation and reversibility as well as the quality of the generated text.
Record the conditions
Track the system version, configuration, data sources, and date. Use consistent criteria for comparisons. Keep uncertainty visible rather than converting every judgement into a precise-looking score. A successful demonstration does not guarantee future behaviour.
Plan ongoing evaluation
Agree how users report failures, who investigates, and what triggers a pause. Revisit the evaluation when providers, prompts, data, permissions, or tasks change. The consequences of failure determine how much scrutiny is appropriate. High-impact use needs particular care.