Three levels. Structural: Markdown exists for every canonical HTML page, identifier parity holds (HTML identifiers == Markdown identifiers, OpenAPI identifiers ⊆ Markdown identifiers), no orphaned descriptions or concatenated card labels, no MDX components leaking into output, exported links resolve without redirecting. Retrieval: canned questions ("should I use SWML or RELAY for a mid-call transfer?", "what replaces the old Agents SDK security page?") must surface the correct product, identifier, and canonical URL. Answer evals: score representative prompts for product routing, parameter accuracy, hallucinated properties, and citations.
Includes a one-time cleanup of redirected URLs still referenced in source content. The parity and redirect-integrity tests double as the regression guards for #525 and #526. Audit "AI quality gates" section and P0 items 3–5.
Three levels. Structural: Markdown exists for every canonical HTML page, identifier parity holds (HTML identifiers == Markdown identifiers, OpenAPI identifiers ⊆ Markdown identifiers), no orphaned descriptions or concatenated card labels, no MDX components leaking into output, exported links resolve without redirecting. Retrieval: canned questions ("should I use SWML or RELAY for a mid-call transfer?", "what replaces the old Agents SDK security page?") must surface the correct product, identifier, and canonical URL. Answer evals: score representative prompts for product routing, parameter accuracy, hallucinated properties, and citations.
Includes a one-time cleanup of redirected URLs still referenced in source content. The parity and redirect-integrity tests double as the regression guards for #525 and #526. Audit "AI quality gates" section and P0 items 3–5.