\begin{table}[htbp]\centering\footnotesize
\caption{Deterministic secondary outcomes by condition (finance held-out documents)}\label{tab:secondary}
\resizebox{\linewidth}{!}{\begin{tabular}{lrrrrrrrrrr}\toprule
Condition & preserved rate & new numbers & n unverifiable citations & n unsupported claims & max overlap share & todo placeholders & n words & cost usd & latency s & n \\\midrule
A\_baseline & 0.05 & 1.00 & 0.08 & 1.08 & 0.07 & 2.54 & 959.33 & 0.14 & 45.88 & 24.00 \\
B\_strong\_prompt & 0.06 & 1.04 & 0.08 & 1.04 & 0.07 & 2.96 & 1027.42 & 0.14 & 46.08 & 24.00 \\
C\_conventional\_rag & 0.05 & 1.25 & 0.08 & 1.62 & 0.07 & 2.17 & 1059.88 & 0.15 & 47.45 & 24.00 \\
D\_profiles\_retrieval & 0.06 & 1.17 & 0.08 & 1.21 & 0.07 & 2.67 & 1051.88 & 0.15 & 49.48 & 24.00 \\
E\_full\_engine & 0.06 & 0.83 & 0.00 & 0.62 & 0.06 & 3.12 & 994.67 & 0.32 & 98.03 & 24.00 \\
E\_no\_factcheck & 0.05 & 0.94 & 0.06 & 0.75 & 0.06 & 2.56 & 1049.94 & 0.31 & 95.92 & 16.00 \\
E\_no\_profile & 0.05 & 0.69 & 0.00 & 0.31 & 0.06 & 2.75 & 996.19 & 0.31 & 97.71 & 16.00 \\
E\_no\_referee & 0.04 & 0.88 & 0.12 & 1.06 & 0.08 & 2.25 & 939.19 & 0.15 & 48.86 & 16.00 \\
E\_no\_retrieval & 0.04 & 0.62 & 0.12 & 0.31 & 0.06 & 5.06 & 1074.38 & 0.31 & 94.53 & 16.00 \\
\bottomrule\end{tabular}}
\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}
\end{table}