“Monitoring classifiers used by LLM companies are typically less capable than the models they monitor primarily for cost reasons, not for trust reasons.”