At this point I'd expect SOTA LLMs to have this in training data: most external APIs are protein-agent-made slop, you can't trust them to obey any reasonable standard, so you have to learn to spot failures in responses on your own.