After an image has been enhanced and has passed the earlier QA (Quality Assurance) checks, it is checked once more before publication. This final decision determines whether the image may be sent to the production menu.
強化中のQAと公開直前QAを分ける
Separate enhancement-stage QA from publish-ready QA
Enhancement-stage QA checks the image during editing. For example, it examines plating, faithfulness to the source image, and colors, and sends feedback back to the editing agent when there is a problem. This is a loop for trying until the image can be improved.
In contrast, publish-ready QA is a gate applied immediately before the enhanced image enters production. It checks not only whether the image violates policy, but also whether it has additional quality problems. While an upstream check focuses on specific items, the final gate looks at the image more holistically.
This diagram shows an enhanced item, a publish-ready gate, policy and quality checks, and the final publication decision. It therefore makes clear that publish-ready QA is not simply a repetition of enhancement-stage QA; it protects against a different risk: publication.
スイスチーズモデルで考える重ねた防御
Think of layered defenses with the Swiss cheese model
The presenters explain this design with the Swiss cheese model. If several slices of hole-filled cheese are stacked, the chance that one path goes through every hole in a straight line becomes smaller. QA works in the same way: specialized checks during enhancement are combined with a broader check immediately before publication. Even when no check is perfect, this lowers the chance that one missed failure reaches production.
Example (created for explanation): Suppose enhancement-stage QA judges the colors and plating acceptable, but publish-ready QA finds a policy problem. In this case, the final gate stops publication. Even if the final check finds a problem that should have been caught upstream, that does not make upstream QA pointless. Layering different viewpoints is itself the intended safety measure.
This system does not promise zero failures. Its purpose is to reduce the chance that a missed issue from any stage enters production. In other words, publish-ready QA is not only a quality-improvement feature; it is also the last defense that checks policy and quality together.