A measure of how decisively the router assigns tokens to experts, indicating healthy expert utilization in MoE models.