1\begin{table}[htbp]\centering\footnotesize2\caption{Judge scores by condition (finance held-out documents, pass A)}\label{tab:judge_means}3\resizebox{\linewidth}{!}{\begin{tabular}{lrrrrrrrr}\toprule4Condition & clarity & argument logic & contribution positioning & methodological rigor & fidelity & usefulness & n invented facts & n \\\midrule5A\_baseline & 5.79 & 5.58 & 5.00 & 4.79 & 5.33 & 5.58 & 1.38 & 24.00 \\6B\_strong\_prompt & 6.04 & 5.71 & 5.33 & 5.04 & 5.33 & 5.71 & 1.62 & 24.00 \\7C\_conventional\_rag & 5.96 & 5.88 & 5.21 & 5.21 & 5.25 & 5.67 & 1.71 & 24.00 \\8D\_profiles\_retrieval & 5.92 & 5.75 & 5.33 & 5.17 & 5.25 & 5.75 & 1.88 & 24.00 \\9E\_full\_engine & 5.83 & 6.04 & 4.92 & 6.12 & 5.71 & 5.42 & 0.88 & 24.00 \\10E\_no\_factcheck & 6.00 & 6.06 & 4.94 & 6.19 & 5.75 & 5.62 & 0.88 & 16.00 \\11E\_no\_profile & 5.94 & 6.12 & 5.12 & 6.38 & 5.69 & 5.69 & 1.00 & 16.00 \\12E\_no\_referee & 6.00 & 5.62 & 4.88 & 4.88 & 5.12 & 5.31 & 1.81 & 16.00 \\13E\_no\_retrieval & 5.88 & 5.94 & 5.19 & 6.19 & 5.62 & 5.56 & 1.50 & 16.00 \\14\bottomrule\end{tabular}}15\begin{minipage}{0.95\linewidth}\vspace{2pt}\scriptsize Notes: generated by scripts/05\_analyze\_benchmark.py from recorded runs (protocol heldout\_v1); judge = claude-sonnet-5-5 blind to condition; scores 1--7.\end{minipage}16\end{table}17