1\begin{table}[t]\centering\small2\caption{103M pilot: held-out cross-entropy (nats/token) and perplexity on 327,680 FineWeb-Edu validation tokens3(40 $\times$ 16 $\times$ 512) under forced random budgets and, for PRISME, under the locked controller.}4\label{tab:pilotppl}5\begin{tabular}{llrr}\toprule6System & policy & loss & ppl\\\midrule7gr 103m & forced b0 & 3.8746 & 48.16\\8gr 103m & forced b1 & 3.8731 & 48.09\\9gr 103m & forced b2 & 3.8722 & 48.05\\10gr 103m & forced b4 & 3.8714 & 48.01\\11gr 103m & forced b8 & 3.8718 & 48.03\\12\bottomrule\end{tabular}\end{table}13