One real run with its thirteen checks, the suite that measures those checks at fourteen out of fourteen, the yield problem a stricter gate created, and one metric cut for failing calibration.
interactive · click and scroll itlive
The mutation table, one row per injected defect with the check that fired, and the one blind spot the suite names instead of hiding.
0 collateralno unrelated check fired on any mutant, and 0 false fails on the clean page.
13 checksdeterministic, rendered at 1440 and 375. The suite needs no API key, so it gates every pull request.
1 blind spotnamed in the results file: invented copy claims, which no deterministic check can catch, handed to the advisory review instead.
What it solves
A model grading its own text output ships pages that look fine in chat and break at 375 pixels. Here the grade comes from rendered checks that run separately from the author model, and the rules live in a file with version history instead of a paragraph in a prompt.
Who it serves
Anyone producing landing pages against a brand system who wants the quality bar automated.
Who it is for
Teams that treat 'looks good in chat' as unproven.
How it works
Brief inThe request arrives with the brand rules it has to satisfy.
Claude drafts to a versioned rules fileThe rules are a file with history, not a paragraph in a prompt.
Render and check, both viewportsThe page is rendered at desktop and mobile and checked against the rules automatically.
The checks decidePass or fail comes from the checks, not the author. Exposed as an MCP server so any client can call it.
Then the checks get testedA suite breaks a page that already passed, one defect at a time, and fails CI if a check stops catching its own defect. It runs with no API key, so it gates every pull request.
Built with
Claude (Sonnet)MCP serverart direction stagemutation evals in CI