The test infrastructure used to run model or agent evaluations repeatedly against examples, expected behavior, tools, and scoring logic. FDEs add real failure cases from production.