Reference implementations generated from a paper's own text, then checked against it — not a link to whatever repo turned up. Each sample carries a verification level on a fixed ladder, and the level shown is never asserted without the report behind it: V0 imports cleanly, V1 executes, V2 two independently-generated implementations agree behaviorally, V3 paper-derived property tests pass on both. A sample never raises the readiness of the paper it comes from.