Auditing black-box clinical prediction models with a large language model through queries alone: a blinded test with planted flaws — ThinkLLM