\section{Conclusion}
\label{sec:conclusion}
We have built an MLX-native MoE runtime whose bounded expert residency is semantically invisible (bit-identical logits to mlx-lm
on two architectures, on one node and across two) and whose all-resident path costs nothing (489 vs 494~\tps{}; 121.6 vs 121.7 on
the medium model), characterised the cluster as it actually is (gigabit Ethernet, no Thunderbolt), and tested the five hypotheses.
Three findings carry over to any similar system: the storage layout of experts decides the cold tier (8 to 36~\tps{}); the number of
host-side operations per installed expert, not prediction accuracy, decides whether prefetching pays (a probe within 0.16 of the
oracle hit rate only wins on long generations); and on gigabit Ethernet a remote expert costs more than a local cold miss, so
placement across nodes buys capacity and, when cost-aware, a better loss than static sharding, but not speed. Kernel fusion was
immaterial because decoding at batch size one is launch-bound. The next version follows from the attribution: a single-copy
install, the probe on the medium model, nodes whose budget the model exceeds, energy on the cluster, and, if cabled, the Thunderbolt
tier. Every number here has a log; the losses are reported with the same care as the wins.