Verified design iteration: Website Console proof

Question

Can Design AI preserve two exact design reviews and explain what changed without turning missing evidence into an improvement claim?

Test subject

The internal Website Console was used as a controlled development subject. The baseline and candidate were canonical quality-report fixtures with matching subject, brief, route, locale, and viewports. This is repository-owned development evidence, not external customer research.

Result

The P13 operation produced a full design-ai-review-comparison v1 artifact. It preserved both references, source strings, parsed reports, byte counts, and SHA-256 digests. The sample remained attention-required because unchanged findings stay persistent; the operation did not manufacture an improvement from equal inputs.

Website Console imported and restored the full artifact, rendered its decisions, and exported bytes with SHA-256 c274093e1ba0f2e9ad21208ae7c512efc0e891f1d15b52cfac7e5a6a94f81b0f, matching the imported file.

Interface evidence

• Desktop viewport: 1440 by 1100, no horizontal overflow.
• Mobile viewport: 390 by 844, no horizontal overflow.
• Minimum visible navigation control height: 44 pixels.
• Keyboard skip link moved focus to the main content.
• Browser console: zero warnings and zero errors in both checks.

Screenshots and the detailed run record are stored under evidence/p13/ in the source repository. They are development evidence and are not included in the npm package.

Permission boundary

The comparison read two explicit report sources and made no repository mutation, network call, deployment, commit, push, or external write. Those actions remained separate gates. The browser check ran only against a local static server.

Claim boundary

Class	Supported statement
Real	The repository implementation, package contracts, and Console import/export path were exercised locally.
Synthetic	The compared review reports were controlled fixtures.
Inferred	The shared contract can support a before-and-after review loop when callers preserve matching context.
Unverified	External adoption, customer finding precision, production quality, business impact, and manual screen-reader output.

This case proves contract behavior and one internal interface path. It does not prove customer adoption or production outcomes.