\begin{tabular}{lrrrrrrrr}
\toprule
Cell & Mean $|P/V-1|$ (\%) & Time to absorb (ticks) & Quoted spread (bp) & Bubble episodes & Max drawdown & LLM order-flow corr. & LLM synchrony & LLM valid-decision rate \\
\midrule
0\% (control) & 0.73 & 36.43 & 18.61$^{***}$ & 0.00 & 0.03 & -- & -- & -- \\
25\% single model/prompt & 0.50 & 28.43 & 10.95 & 0.00 & 0.02 & 0.40 & 0.95 & 0.98 \\
25\% diverse prompts & 0.53$^{**}$ & 24.00$^{**}$ & 11.47 & 0.00 & 0.03$^{*}$ & 0.32$^{**}$ & 0.83$^{***}$ & 0.97$^{**}$ \\
25\% two models & 0.50 & 24.00$^{**}$ & 15.57$^{***}$ & 0.00 & 0.03$^{***}$ & 0.29$^{**}$ & 0.83$^{***}$ & 0.98 \\
25\% memory (5) & 0.54$^{***}$ & 31.14 & 10.71 & 0.00 & 0.03$^{***}$ & 0.44 & 0.95 & 0.99$^{*}$ \\
25\% communication channel & 0.55$^{***}$ & 28.00 & 11.63 & 0.00 & 0.02 & 0.48$^{**}$ & 0.93 & 0.98 \\
25\% adaptive feedback & 0.50 & 26.57 & 10.62 & 0.00 & 0.02 & 0.41 & 0.95 & 0.98 \\
0\% + leveraged 19\% (classical) & 2.16$^{***}$ & 53.00 & 38.82$^{**}$ & 0.71$^{*}$ & 0.07$^{***}$ & -- & -- & -- \\
25\% + leveraged 19\% & 0.96$^{***}$ & 16.29$^{**}$ & 38.51$^{***}$ & 0.00 & 0.06$^{***}$ & 0.46 & 0.95 & 0.98 \\
25\% + circuit breaker & 0.50 & 28.43 & 10.95 & 0.00 & 0.02 & 0.40 & 0.95 & 0.98 \\
25\% + position limit & 0.51 & 34.00 & 22.48$^{***}$ & 0.00 & 0.02 & 0.39 & 0.94 & 0.96$^{***}$ \\
5\% gpt-5.4-mini (equal USD) & 0.64$^{**}$ & 19.57$^{**}$ & 22.20$^{***}$ & 0.00 & 0.03$^{*}$ & 0.52$^{**}$ & 0.98$^{**}$ & 0.99 \\
25\% matched fast latency & 0.51 & 29.57 & 12.09 & 0.00 & 0.03$^{**}$ & 0.43 & 0.95 & 0.98 \\
\midrule
Seeds per cell & 7 & & & & & & & \\
\bottomrule
\end{tabular}