The model writes the page. Separate checks decide if it ships.
One real run for a client brief, exactly as the engine logged it: the rendered page at both viewports, the deterministic checks, and the model's advisory review.
what it rendered · desktop 1440 and mobile 375
page-1440.png
page-375.png
deterministic checks · these decide the verdict
| result | check | rule |
|---|---|---|
| pass | document structure | valid-document |
| pass | single file, no external assets | single-file |
| pass | link audit (0 external links) | cta-specific |
| pass | no horizontal overflow at 375px | responsive |
| pass | axe-core, no serious or critical violations (30 rules passed) | semantic-html / contrast-aa |
| pass | copy tells (dashes, banned vocabulary) | copy-tells |
verdict: pass
drafted by claude-sonnet-5
cost $0.1973
model review · advisory only, it does not decide
cta-specific (minor): the brief's phone number is +63 2 8845 2210, eleven digits, but the tel: link drops a digit, so the href does not match the real contact destination from the brief.
href="tel:+6328845220"
The split is the point. Six deterministic checks decide pass or fail; the model reviews alongside and can only advise. Here every check passed, and the advisory pass still caught a real defect a human would have shipped: a phone number missing one digit. Verbatim from the run's own qa-report, 2026-07-15.