Build a Source-Grounded Answer Test for Your Knowledge Twin

ROLE Act as a knowledge-assistant evaluation designer. OBJECTIVE Build a benchmark from [KNOWLEDGE MATERIALS] for [USER TASKS], using [FAILURE PRIORITIES] to determine which mistakes matter most. INSTRUCTIONS Create fifteen test questions covering direct lookup, synthesis across sources, ambiguous requests, outdated information, conflicting guidance, and questions the materials cannot answer. Include realistic wording that users would actually type rather than only document headings. For each question, specify the expected answer elements, supporting source passages or locators, unacceptable claims, and whether clarification or abstention is required. Separate questions whose answers are documented from hypothetical stress tests. Define a practical scoring rubric for factual support, relevance, completeness, citation fidelity, and boundary handling. Recommend a repeatable test procedure with a baseline run, failure review, targeted changes, and regression testing. Explain how to keep some questions out of the assistant's setup instructions to reduce teaching directly to the test. CONSTRAINTS Avoid fabricated statistics and flag assumptions. Do not invent citations or fill source gaps with general knowledge. OUTPUT FORMAT Provide the benchmark table, scoring rubric, critical-failure rules, and test procedure.