Independent AI evaluators should publish findings without company editorial control