% Generated by paper_plot_script/make_figures_tex.py -- do not edit.
% The caption text and the table numbers live in that chart's plot script;
% re-run the script to update this file.
\begin{table}[!hb]
\centering
\small
\caption{\textbf{Manager vs.\ single call, five passes.} pass@1 on LCB-100 at a 128k cap, reasoning on; $\pm$ is a 95\% $t$ interval across the five passes (df = 4). $\Delta$ is in percentage points, against the model's own single call and against Fable 5's single call. Each $p$ is a paired sign-flip permutation test, unit = problem ($n = 100$), Holm-corrected within its family of five; $<$ marks the permutation floor. \textbf{95\% bound} is the one-sided lower bound on $\Delta$ vs Fable 5 over the same per-problem differences (t, df = 99): the largest deficit the data leave open.}
\label{tab:main}
\begin{tabular}{@{}l@{\hspace{0.55em}}c@{\hspace{0.55em}}c@{\hspace{0.55em}}r@{\hspace{0.55em}}c@{\hspace{0.55em}}r@{\hspace{0.55em}}r@{\hspace{0.55em}}c@{}}
\toprule
\textbf{Model} & \textbf{\shortstack{Single\\call}} & \textbf{\shortstack{With\\manager}} & \textbf{\shortstack{$\Delta$ vs\\single}} & \textbf{$p$} & \textbf{\shortstack{$\Delta$ vs\\Fable 5}} & \textbf{\shortstack{95\%\\bound}} & \textbf{$p$} \\
\midrule
Claude Fable 5 & $90.4 \pm 1.4$ & --- & --- & --- & reference & --- & --- \\
GPT-5.6-Terra & $80.8 \pm 1.0$ & $88.0 \pm 1.2$ & $+7.2$ & $8.0\times10^{-5}$ & $-2.4$ & $-5.8$ & $0.88$ \\
GPT-5.6-Luna & $70.4 \pm 4.6$ & $81.2 \pm 2.8$ & $+10.8$ & $<2.5\times10^{-5}$ & $-9.2$ & $-13.4$ & $0.0013$ \\
Qwen3.8-27B & $66.8 \pm 5.2$ & $92.4 \pm 2.3$ & $+25.6$ & $<2.5\times10^{-5}$ & $+2.0$ & $-0.8$ & $0.88$ \\
Qwen3.8-Flash-Next & $84.2 \pm 1.8$ & $93.0 \pm 1.2$ & $+8.8$ & $<2.5\times10^{-5}$ & $+2.6$ & $-0.1$ & $0.56$ \\
DeepSeek-V4.1-Flash & $88.4 \pm 2.6$ & $92.0 \pm 1.8$ & $+3.6$ & $0.002$ & $+1.6$ & $-1.3$ & $0.88$ \\
\bottomrule
\end{tabular}
\end{table}