Freeze the test conditions
Record provider, product, model label, region, date, inputs, controls, prompt, attempt budget, reviewer, and delivery target. Do not pool results from different versions or interfaces.
How to work through it
Score observable criteria
Review instruction following, identity, geometry, motion, camera, continuity, text, physical behavior, failure rate, and required repair. Use the same rubric for every candidate.
Make a bounded decision
Report sample size, selected and rejected attempts, editorial time, unresolved risk, rights checks, and export fit. Limit the conclusion to the tested setup and date.