\begin{table}[t]\centering\small
\caption{Systems and controls. ``bb'' = backbone, ``$6NP$'' = the usual training-FLOP estimate for $N$ tokens and $P$
parameters; adapter training additionally executes up to 8 editor applications per position. Adapter parameters are
unique parameters outside the backbone (counted by instantiation; the 103M editor includes the residual-plus-entropy
controller features).}
\label{tab:controls}
\begin{tabular}{lrrlll}\toprule
System & backbone & adapter & training exposure & training compute & inference regime\\\midrule
Frozen backbone v2 (budget 0) & 4,754,688 & 182,725 & synthetic train, 5k steps $\times$ 64 & backbone only & no workspace (bitwise bypass)\\
PRISME sparse editor (tiny) & 4,754,688 & 182,725 & same + 3k adapter steps $\times$ 32 & frozen bb fwd + editor & 1--2 slots/action, 0--8 actions\\
Global-recurrence adapter (tiny) & 4,754,688 & 172,608 & same & frozen bb fwd + block & whole latent/action, 0--8 actions\\
Dense 16-slot editor (tiny) & 4,754,688 & 182,725 & same & frozen bb fwd + 16 lanes & 16 slots/action, 0--8 actions\\
PRISME-103M sparse editor & 99,502,848 & 2,762,757 & FineWeb-Edu 98.3M tok + 24.6M adapter tok & $6NP$ + editor & 1--2 slots/action\\
Global-recurrence adapter (103M) & 99,502,848 & 2,623,488 & same & $6NP$ + block & whole latent/action\\
Plain decoder, parameter-matched (102.2M) & 102,157,056 & 0 & FineWeb-Edu 98.3M tok (same recipe) & $6NP$ & no workspace\\
\bottomrule\end{tabular}\end{table}