●builderYou must match your inference-time expert pruning strategy to your model's training-time routing logic to avoid massive perplexity spikes.
●researcherThis establishes a formal weight-sum model to explain how architectural changes during inference interact with training objectives.