>> The model was perfectly capable of verifying its correctness.

It's an LLM. It can't do that.