\begin{table}[t]\centering\small \caption{103M pilot: held-out cross-entropy (nats/token) and perplexity on 327,680 FineWeb-Edu validation tokens (40 $\times$ 16 $\times$ 512) under forced random budgets and, for PRISME, under the locked controller.} \label{tab:pilotppl} \begin{tabular}{llrr}\toprule System & policy & loss & ppl\\\midrule gr 103m & forced b0 & 3.8746 & 48.16\\ gr 103m & forced b1 & 3.8731 & 48.09\\ gr 103m & forced b2 & 3.8722 & 48.05\\ gr 103m & forced b4 & 3.8714 & 48.01\\ gr 103m & forced b8 & 3.8718 & 48.03\\ \bottomrule\end{tabular}\end{table}