Public (unauthenticated) storefront endpoint that surfaces the latest judged vendor-bot feature-reliability eval. DOUBLE-GATED and fail-closed: it returns data ONLY when the API runs in the sandbox environment AND {slug} matches the single configured demo vendor (vendor_eval.demo_vendor_slug). Any other slug, any other environment, or an unset config returns a 404 indistinguishable from a missing route.
The eval suite is vendor-agnostic (it seeds its own throwaway world), so the slug is a VISIBILITY gate, not a data filter: the payload is the whole suite's latest scored run. Only a whitelisted set of top-level scores keys is ever re-served (headline, matrix, features, scenarios, top_failures, flaky_scenarios, feature_order, axis_order, feature_labels, axis_labels, generated_at); anything else the judge stores stays server-side.
Rate-limited to 30 requests per minute per client (429 beyond that).
All rate metrics are fractions in [0, 1] (0.84 = 84%). pass_rate is the share of samples that passed; pass_k is the stricter share of scenarios that passed ALL K samples (only differs from pass_rate when samples > 1).