Turn a PRD and design into a reviewable Surface Contract
A3S Test uses a deployment-injected provider to interpret PRD spans and design regions as precisely sourced candidate structure. A person must review every candidate, resolve conflicts, and close open decisions before publishing a Surface Contract. At runtime, A3S Test reconciles that contract with the browser accessibility tree and Test Kit Page Context, retaining provenance, current page revision, and evidence for every difference.
An accessibility tree is a browser fact computed from real DOM, native control semantics, and current state. PRDs and designs describe expectations but cannot prove that a page implements those semantics. A Surface Contract preserves expectations, provenance, and uncertainty for comparison with browser facts instead of presenting model inference as page observation.
Separate expectations from facts
Only a reviewed selection can become a blocking Surface Contract. Model output, unresolved conflicts, and low-confidence candidates remain advisory.
Generate candidates
An ACL file declares sources, budgets, and the deployment-owned provider:
The draft contains candidates, conflicts, open decisions, provider usage, and source digests. It has no executable authority.
Review and publish
Review rehashes sources, checks cited byte ranges, rejects unresolved conflicts or product questions, and regenerates canonical ACL locally. The workflow checksum detects accidental or unreviewed mutation; it is not a digital signature.
Reconcile against the rendered page
A suite references the reviewed contract with verify_contract:
Reconciliation prefers test ID, component identity, role plus accessible name, and finally role. Missing or truncated Test Kit context is inconclusive and fails closed. Blocking differences determine the suite result. Advisory differences remain in the report.
Optional visual grounding and design audit
A visual-grounding provider receives an observation- and SHA-256-bound PNG and returns point or box candidates. A3S Test rehashes the bytes, admits geometry, and hit-tests current Test Kit nodes. A unique hit may become a current semantic target. Ambiguous or unmapped results remain image-bound advice. Grounding never clicks automatically.
A design-audit provider receives the screenshot plus complete page context and may advise on hierarchy, composition, spacing, typography, color use, consistency, interaction clarity, content clarity, and responsive composition. Advice has no verdict, action, or repair authority and reaches the Repair Ledger only after human review.
