Keep evaluation from becoming rehearsal
If developers repeatedly tune the prompt against the same cases and then report performance on those cases, the result may overstate performance on new work. Separate development material from evaluation material and manage access to reference answers.
Worked study design — fictional
A team collects representative fictional protocols and requirements. Independent qualified reviewers create reference findings using a defined rubric. Disagreements are adjudicated. Development cases help improve the application. A separate held-out set is used for the planned evaluation after configuration is fixed.
The split should consider near duplicates, shared templates, related projects, and repeated passages. Moving two nearly identical documents into different sets does not necessarily create meaningful independence.
Reference quality matters
A reference answer can also be wrong. Record reviewer qualifications, instructions, issue definitions, disagreements, and limitations. For ambiguous questions, an acceptable response may be to request clarification instead of providing one confident answer.
Exercise
The team discovers a failure on the held-out set, fixes the prompt, and reruns the same set. Can it report the rerun as untouched independent evaluation?
Show suggested answer
No. The set has informed development. Report that history honestly and use an appropriate fresh evaluation or justified design for the revised claim. Preserve the original results and the reason for the change.
Expected records
A reference-data specification, independent scoring rubric, adjudication log, protected set manifest and leakage assessment. Adapt record names and detail to the applicable procedures; these are practitioner examples, not a universal mandatory document list.
Relevant regulations and guidance
Regulations set requirements within their scope; guidance offers recommendations or explains applicable expectations; standards depend on adoption or incorporation. Voluntary frameworks do not establish compliance. The lesson's controls, test designs and example records are practitioner suggestions unless explicitly identified otherwise.
Voluntary framework · National Institute of Standards and Technology · NIST AI 100-1 · January 2023
Scope: United States · Cross-sector AI governance and risk management, including generative AI.
Relevance to this lesson: Supports documenting test sets and deployment-relevant evaluation conditions.
Publication status: Final / published. Source checked 21 September 2026.
Publication identity, scope and status checked against the linked official source. This is not an independent audit of every retained guide statement.
Source details and applicabilityGuidance · European Medicines Agency · EMA/INS/GCP/112288/2023
Scope: European Union · Sponsors, investigators and service providers using computerised systems and electronic data in clinical trials.
Relevance to this lesson: For clinical-trial systems, consider trial-specific data and system responsibilities; the fictional protocol exercise does not establish clinical-trial suitability.
Publication status: Final / published. Source checked 21 September 2026.
Publication identity, scope and status checked against the linked official source. This is not an independent audit of every retained guide statement.
Source details and applicability