A team wants to know whether a prompt revision improved its application. Which evaluation…

A team wants to know whether a prompt revision improved its application. Which evaluation design gives the clearest evidence?

1. Run the old and new prompt against the same versioned cases, model settings, rubric, and aggregation method
2. Change the prompt, model, retrieval corpus, and rubric together, then compare one aggregate score
3. Ask the prompt author which version feels better without recording outputs