Design DNA

Coding agents ship UI that looks finished and quietly drifts — a radius that doesn't match, spacing that wanders, two buttons that disagree. Nobody catches it, because checking means inspecting forty elements by hand.

Design DNA loads the page in a real headless browser, scrolls it so reveal animations actually fire, and measures every rendered element in the same pass it screenshots — so every finding can be pointed at on the page.

Findings come back in four classes: PASS, FAIL, REVIEW, UNSUPPORTED. REVIEW is the important one — a measured difference that might be intentional is never auto-merged. On the demo page it correctly declines to flag a deliberately distinct CTA.

Corrections are merge-only: it snaps a stray value to one already on the page and never invents a new one. Each correction is emitted as an executable postcondition — a property, target element ids, and an expected value. A re-scan passes only when every named target resolves to expected. Never a page-wide count, which would hide a change to the wrong element.

No model is used anywhere in the product. Same page in, same numbers out. The AI in this story is Cursor, doing the repair from the evidence.

On stripe.com: 2,722 elements, 17 tiles, a 15,182px page, 9 contract failures.

public/corpus.json is a dataset of 573 sites fingerprinted on 18 Sep 2026, before this event. It is data, not code. All code in this repository was written during the hackathon.