\begin{table}[t]\centering\small
\caption{Phase 4 decisive comparison on the frozen 5.6M-parameter synthetic backbone (char-level): held-out
answer-span loss on dev-iid under random forced actions at fixed budgets, and free-generation exact match
(dev-iid, 1,000 stratified items). Mean $\pm$ SD over independent training seeds of the adapter; the backbone,
data, recipe and symmetry-breaking device are identical. Budget 0 is the frozen backbone for every system.}
\label{tab:seedmatrix}
\begin{tabular}{llrrrrr}\toprule
System & seeds & loss b=0 & loss b=1 & loss b=2 & loss b=4 & loss b=8\\\midrule
PRISME sparse editor & 4 & 0.4621 $\pm$ 0.0000 & 0.4526 $\pm$ 0.0006 & 0.4493 $\pm$ 0.0007 & 0.4483 $\pm$ 0.0017 & 0.4573 $\pm$ 0.0035\\
Global recurrence (inspired) & 4 & 0.4621 $\pm$ 0.0000 & 0.4452 $\pm$ 0.0014 & 0.4430 $\pm$ 0.0026 & 0.4443 $\pm$ 0.0036 & 0.4456 $\pm$ 0.0038\\
\midrule
\multicolumn{7}{l}{\emph{Exact match (\%), dev-iid}: b=0 / b=1 / b=2 / b=8}\\
PRISME sparse editor & 4 & \multicolumn{5}{l}{37.7 $\pm$ 0.0 / 39.2 $\pm$ 0.1 / 39.5 $\pm$ 0.2 / 39.2 $\pm$ 0.5}\\
Global recurrence (inspired) & 4 & \multicolumn{5}{l}{37.7 $\pm$ 0.0 / 39.8 $\pm$ 0.3 / 39.6 $\pm$ 0.4 / 40.0 $\pm$ 1.3}\\
\multicolumn{7}{l}{\emph{Exact match (\%), held-out template / depth / compose at b=2}}\\
PRISME sparse editor & 4 & \multicolumn{5}{l}{27.5 $\pm$ 0.5 / 29.4 $\pm$ 0.2 / 3.1 $\pm$ 0.3}\\
Global recurrence (inspired) & 4 & \multicolumn{5}{l}{27.7 $\pm$ 0.2 / 29.9 $\pm$ 0.3 / 2.8 $\pm$ 0.3}\\
\bottomrule\end{tabular}\end{table}