1\begin{table}[t]2\centering\small3\caption{Trace-driven simulation (OLMoE, per-layer LRU tables, cold per prompt): hit rate and modelled stall per token by predictor, split and capacity $C$.}4\label{tab:sim}5\begin{tabular}{llrrrrrrrr}6\toprule7Predictor & split & \multicolumn{2}{c}{$C=16$} & \multicolumn{2}{c}{$C=24$} & \multicolumn{2}{c}{$C=32$} & \multicolumn{2}{c}{$C=48$} \\8 & & hit & stall & hit & stall & hit & stall & hit & stall \\9\midrule10none & fit & 0.577 & 23.3 & 0.706 & 16.2 & 0.804 & 10.8 & 0.931 & 3.8 \\11none & heldout & 0.581 & 23.1 & 0.706 & 16.2 & 0.806 & 10.7 & 0.933 & 3.7 \\12prev\_rank & fit & 0.545 & 25.0 & 0.665 & 18.4 & 0.770 & 12.7 & 0.919 & 4.5 \\13prev\_rank & heldout & 0.556 & 24.4 & 0.670 & 18.1 & 0.771 & 12.6 & 0.920 & 4.4 \\14transition & fit & 0.680 & 17.3 & 0.770 & 12.5 & 0.839 & 8.8 & 0.940 & 3.3 \\15transition & heldout & 0.686 & 17.0 & 0.771 & 12.5 & 0.840 & 8.7 & 0.942 & 3.2 \\16transition2 & fit & 0.663 & 18.0 & 0.758 & 13.0 & 0.832 & 9.1 & 0.938 & 3.4 \\17transition2 & heldout & 0.672 & 17.6 & 0.761 & 12.9 & 0.833 & 9.0 & 0.940 & 3.3 \\18transition3 & fit & 0.651 & 18.5 & 0.749 & 13.5 & 0.826 & 9.4 & 0.937 & 3.4 \\19transition3 & heldout & 0.660 & 18.1 & 0.752 & 13.3 & 0.828 & 9.3 & 0.939 & 3.3 \\20oracle & fit & 0.938 & 3.0 & 0.959 & 2.0 & 0.973 & 1.4 & 0.991 & 0.5 \\21oracle & heldout & 0.940 & 2.9 & 0.959 & 2.0 & 0.972 & 1.4 & 0.991 & 0.5 \\22probe & fit & 0.813 & 9.8 & 0.868 & 7.0 & 0.908 & 4.9 & 0.965 & 1.9 \\23probe & heldout & 0.811 & 10.0 & 0.864 & 7.3 & 0.904 & 5.1 & 0.965 & 1.9 \\24\bottomrule25\end{tabular}26\begin{tablenotes}\footnotesize27\item Source: \code{benchmarks/processed/gateC/sim\_olmoe\_v2\_ssd043\_m8.json}; t\_read 0.43 ms, t\_layer 0.128 ms, 1 loader(s); hit = per-layer hit rate, stall in ms per token (modelled).28\end{tablenotes}29\end{table}30