Text this: GPT-4.1 and Llama 3.3 70 fail to detect clinically relevant errors in radiology reports in zero-shot evaluation.