Aren't they running a safety classifier model in parallel to the main user thread to catch and prevent 'banned' outputs?