1\section{Limitations}2\label{sec:limitations}3\begin{itemize}4 \item \textbf{Scope of evidence.} Version 0.1 reports single-node results on a small model whose experts fit in memory, so the5 regimes where H1--H3 can matter (expert set larger than the budget of a node) are only approached by artificially bounding the6 slot tables. The medium model on the 36--64~GB nodes is the first realistic regime.7 \item \textbf{Topology.} The cluster is connected by gigabit Ethernet with no Thunderbolt link; RDMA/JACCL results would require8 cabling the Studios (a physical change outside the agent's authority). A LAN fetch of one expert costs about 28~ms, which rules9 out moving weights during decode and confines the remote tier to remote \emph{execution}.10 \item \textbf{Synchronisation floor.} Any demand-loading design must read the router's choice on the host once per layer; on this11 model that costs 2.4$\times$ in throughput. Hiding it requires either complete tables (no benefit from caching) or speculative12 execution on predicted experts with verification, which changes nothing semantically only if the verification is exact --13 a Gate D question.14 \item \textbf{Predictors.} Only simple predictors are tested; the linear probe is under-trained with the small trace corpus.15 \item \textbf{Measurements.} Stall times in simulation use point estimates of tier latencies; online runs use the page-cache tier16 unless stated; energy is not yet measured.17 \item \textbf{Baselines.} llama.cpp and colibri comparisons are pending and will carry precision flags because their18 quantization formats differ from MLX's.19\end{itemize}20