Your AI builder finished the site.
That doesn't mean the design is finished.
Generation is not judgment.
Website QC reads the code and the rendered page together, then reports the specific defects the generator left behind. Dead framework classes, fragile external assets, buttons whose copy does not match their handler, missing structural sections, duplicated components. Checkable things, not taste.
That is the product in one gesture. A generator will happily produce six sections of identical weight, two of which are the same component pasted twice, because nothing in it is asking what the page is for.
The same block, six times, at the same visual weight. Common in generated output because each section is produced independently and nothing reconciles them at the end.
Components that exist twice, framework classes referenced by nothing, imports for code that was deleted three prompts ago. Dead weight a reader has to step over.
Template assumptions nobody chose. Placeholder copy that shipped, a pricing section on a site with no pricing, hero images pointed at a CDN that will not be there next year.
A button that says one thing and does another. A form with no failure state. The gap between what the interface promises and what the code behind it handles.
Manual structural review. You send a repository or a URL. It gets read by hand, with tooling, and you get back a written report of specific defects with the reasoning for each. This is a service, delivered by a person. It is what exists today.
Repository-aware reviewer. The direction is a reviewer that reads the repo itself and proposes small, verifiable patches as pull requests. That does not exist yet. It is not sold, not scheduled, and nothing above depends on it.
Tested against real repositories, not demos.
Two validation passes are done, against public repositories built with real AI builders: Lovable family output in the first, a Bolt.new project in the second. Both were cloned, installed, run locally and patched in place rather than reviewed from a screenshot.
To be exact about the limits: the evidence so far is one snapshot per repository, and auth-gated internal screens were not reviewed. Continuous monitoring is not something this has demonstrated, so it is not claimed.