Fix the task first
Use the same inputs, model, and settings when comparing versions. Write the success criteria before seeing either result. Otherwise it is easy to reward whichever answer happens to sound more confident.
Build a tiny evaluation set
Include a representative case, a messy case, a missing-information case, and an input that could tempt the model to invent facts. Use examples you have permission to test. Keep personal data out of the evaluation set.
Score the work, not the style
Assess factual accuracy, coverage, grounding, and usefulness separately. Hide which prompt produced which output if possible. Record failures even when most of the answer looks good.
Repeat and record
A single output is not proof of reliability. Repeat important cases within your available budget. Record dates and versions, and describe the limitations of your sample alongside any recommendation.
Find your starting point