Sibling contract
Drift
AI-generated UI can drift off the tokens a site declares. This contract names four drift modes and defines the checks that detect drift deterministically from compiled CSS rather than prose advice.
Version 0.1.1 · provisional · machine export
Source authority
- Scale of the problem
- ~160 visual issues per AI-generated app
- OverlayQA · Jason Arbon, 1000+ checks
- Primary source
- Designesy design intelligence research 2026-08-01
The four drift modes
Conformance
Detection method
Fetch the URL HTML + all linked CSS. Extract :root custom property declarations and all var() references. Compare declared tokens against referenced tokens. Measure value variance across declarations for spacing, color, radius, shadow, and timing.
Verification
12 checks. PASS=1, WARN=0.5, FAIL=0. Score = (points/12) × 100. A≥90, B≥80, C≥70, D≥60, F<60.
Validation
Designesy drift engine: CSS extraction + var() reference analysis. Method: Fetch HTML + linked CSS, extract :root declarations and var() calls, compute variance. Browser-only checks: None: all 12 checks are static CSS/HTML analysis, no browser needed.
Open questions
- Between-session amnesia (mode 3) requires historical snapshots: the current engine detects within-page drift only
- Silent breaking changes (mode 4) require version tracking: the current engine detects the current state only
- Cross-page drift (mode 2) requires crawling multiple pages: the current engine analyzes the single fetched URL
- The engine may flag third-party CSS (analytics, embeds) as fabricated tokens: a suppression list may be needed
Run it
Score any URL for drift: