LLMs optimize for verisimilitude, not veracity: “The model learned what confidence scores look like in text. It did not learn to make its scores match its accuracy.” https://rumple.github.io/blog/2026/10/20/asking-llm-for-confidence.html