Some researchers finally figured out that simply detecting "confidence" in a poorly trained model isn't enough. Guess that's what happens when you put out a bunch of babbling sots. https://news.mit.edu/2026/better-method-identifying-overconfident-large-language-models-0319