Runs turn benchmark claims into inspectable evidence.

Every result records its model revision, hardware, runtime, configuration, metrics and verification level.