\begin{table}[htbp]\centering\footnotesize
\caption{Deterministic secondary outcomes by condition (finance held-out documents)}\label{tab:secondary}
\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 & 0.20 & 1.58 & 0.00 & 1.17 & 0.04 & 12.29 & 769.33 & 0.08 & 33.67 & 24.00 \\
B\_strong\_prompt & 0.28 & 1.21 & 0.00 & 1.79 & 0.04 & 15.46 & 828.42 & 0.08 & 36.12 & 24.00 \\
C\_conventional\_rag & 0.24 & 1.46 & 0.00 & 1.50 & 0.03 & 12.62 & 844.62 & 0.09 & 37.61 & 24.00 \\
D\_profiles\_retrieval & 0.24 & 1.83 & 0.00 & 1.46 & 0.03 & 12.46 & 847.50 & 0.10 & 38.20 & 24.00 \\
E\_full\_engine & 0.30 & 0.71 & 0.00 & 0.88 & 0.05 & 12.25 & 981.25 & 0.20 & 80.06 & 24.00 \\
E\_no\_factcheck & 0.40 & 1.94 & 0.00 & 1.00 & 0.03 & 12.88 & 909.06 & 0.20 & 77.61 & 16.00 \\
E\_no\_profile & 0.36 & 0.62 & 0.00 & 1.12 & 0.03 & 13.06 & 931.50 & 0.19 & 77.19 & 16.00 \\
E\_no\_referee & 0.22 & 1.44 & 0.00 & 0.94 & 0.04 & 12.12 & 790.81 & 0.10 & 39.50 & 16.00 \\
E\_no\_retrieval & 0.30 & 0.94 & 0.00 & 0.75 & 0.05 & 14.31 & 897.94 & 0.20 & 78.02 & 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}