Define the task boundary
Specify whether the system summarises evidence, extracts data, compares technologies or drafts explanations. Name the intended reader and the decision supported. A broad benchmark score may conceal weaknesses in the narrow domain where the output will actually be used.
Select representative and difficult cases
Include ordinary questions, ambiguous inputs, conflicting evidence and cases where information is insufficient. Use examples across the relevant streams and environmental conditions. Keep the evaluation set separate from development examples where independence matters, and document how cases were chosen.
Agree the scoring rubric
Evaluate factual accuracy, source support, reasoning, completeness and communication of uncertainty separately. Define errors with examples and distinguish a minor wording issue from a decision-changing mistake. Have domain reviewers discuss disagreements instead of hiding them in a single average score.
Make results reproducible
Record the system version, instructions, inputs and assessment date. Report case-level findings alongside the summary and repeat relevant checks when the workflow changes. A useful evaluation produces an error taxonomy and improvement priorities as well as a score.
Sources & context
Background reading checked 9 October 2026. The review questions and structures above are BGF editorial guidance; the references are not BGF project evidence or endorsements. Apply them to the relevant site, data and agreed scope.