A warning needs to identify the defect
An ad review that says 'check the claims' leaves the work with the reader. A useful review points to the claim, explains the problem and makes the consequence clear. The benchmark rewards matching the actual defect set, so vague caution cannot earn full credit.
The twenty-asset pack includes both images and videos. Some videos contain more than one known issue, and the score reflects how much of that set the model finds. Finding one problem is useful; it is not the same as completing the inspection.
A longer checklist does not guarantee better detection
The baseline asks for launch QA without a checklist, while the scaffold adds one. In the reviewed Claude Fable 5.1 comparison, the mean falls from 6.38 to 5.54 across twenty paired observations. The right question is whether the checklist improves detection on your assets.
Keep the review focused on the dimensions that matter for the creative: product fidelity, visible or spoken claims, copy, offer and format. Ask the model to distinguish a confirmed issue from a question requiring more context, then inspect the evidence it gives.
Use AI to make human review more targeted
A practical handover is a short list of issues with the affected image region or video moment, the reason each matters and the next action. That makes the model's finding easy to verify and route to the person who can fix it.
This pack measures defect coverage on assets with known issues. To test a release gate in your own process, include clean assets too: the cost of flagging a good ad matters alongside the cost of missing a bad one.
How Benmark tests this
Models review ten images and ten videos containing known defects. They are asked to check launch readiness without being told that a defect is guaranteed. Scoring measures how much of the answer-key defect set they find.
Test inputs, instructions and scoring
What the model receives
Both prompt conditions receive the same twenty media assets. The evaluator's key records the planted issue and observed secondary video defects. A complete run has forty reviews: twenty assets under each condition.
What changes between prompts
The simple prompt asks for launch QA; the scaffold adds a content-review checklist. The checklist directs the inspection without telling the model which defect is present in an asset.
How the result is scored
A single-defect asset scores 10 when the expected defect is found, 5 when an issue is flagged without matching that defect, and 0 when it is missed. Multi-defect video scores reflect the share of expected defects matched. Semantic interpretation resolves eligible answer matches. The pack does not establish a clean-ad false-alarm rate.