New design contract benchmark¶
Source¶
This is a synthetic regression case, not a customer project or an adoption claim.
The suite brief asks for a Korean team workspace for reviewing pending approvals
without losing context. The inspected candidate is the packaged
candidate.html
fixture.
Change¶
design-ai start turns the brief into the existing design-from-brief route and
an agent-readable DESIGN.md contract. The benchmark does not generate or edit a
product screen. It checks that the first artifact still names its route, output,
workflow, verification, and approval boundary before implementation begins.
Verification¶
Run:
The case passes only when the start payload and embedded design contract keep their versioned identities, the review remains unexecuted, every performed write array is empty, and target-repository editing still requires approval.
Permission boundary¶
The case reads the shipped corpus to compose a plan. It writes no local file, does not inspect or mutate a target repository, and makes no external request. The next implementation action remains behind the human approval list in the artifact.
Remaining risk¶
The benchmark proves contract composition, not visual quality or first-value time. No interface, runtime, accessibility tree, responsive layout, or user outcome was observed. Those claims remain unverified until a real implementation is reviewed.
Claim boundary¶
This synthetic case does not prove customer adoption, production quality, or the five-minute first-value target. It proves only that the local design contract and fixture inspection retain their declared evidence and approval boundaries.