1\begin{table}[htbp]\centering\footnotesize2\caption{LLM laboratory (CDA): policies $\times$ signal structure}\label{tab:llm}3\begin{tabular}{llrrrrrrr}\toprule4Policy & $(D,M)$, temp. & corr.\ $|$ public & raw corr. & vol.\ (\%) & trades & invalid & clipped & s/run\\\midrule5Bayesian control & (0,1), 0.0 & 0.12 (0.07) & 0.35 & 0.17 & 6.9 & 0 & 81 & 0\\6Bayesian control & (1,1), 0.0 & 0.65 (0.08) & 0.73 & 0.15 & 4.4 & 0 & 71 & 0\\7Bayesian control & (1,3), 0.0 & 0.47 (0.05) & 0.53 & 0.18 & 6.0 & 0 & 64 & 0\\8three LLMs (Qwen/Mistral/Llama) & (0,1), 0.0 & 0.07 (0.03) & 0.08 & 0.56 & 5.2 & 0 & 120 & 185\\9three LLMs (Qwen/Mistral/Llama) & (0,1), 0.7 & 0.07 (0.09) & 0.09 & 0.99 & 5.6 & 0 & 149 & 202\\10three LLMs (Qwen/Mistral/Llama) & (1,1), 0.0 & 0.16 (0.07) & 0.14 & 0.27 & 5.6 & 1 & 109 & 196\\11three LLMs (Qwen/Mistral/Llama) & (1,1), 0.7 & 0.17 (0.08) & 0.18 & 0.25 & 6.1 & 2 & 118 & 170\\12three LLMs (Qwen/Mistral/Llama) & (1,3), 0.0 & 0.08 (0.03) & 0.09 & 0.80 & 5.9 & 1 & 107 & 194\\13three LLMs (Qwen/Mistral/Llama) & (1,3), 0.7 & 0.18 (0.05) & 0.16 & 1.25 & 5.8 & 2 & 141 & 135\\14one LLM (Qwen2.5-7B) & (0,1), 0.0 & 0.12 (0.02) & 0.20 & 0.32 & 10.9 & 0 & 90 & 147\\15one LLM (Qwen2.5-7B) & (0,1), 0.7 & 0.20 (0.09) & 0.29 & 0.35 & 10.7 & 0 & 85 & 147\\16one LLM (Qwen2.5-7B) & (1,1), 0.0 & 0.20 (0.10) & 0.23 & 0.30 & 11.1 & 0 & 87 & 143\\17one LLM (Qwen2.5-7B) & (1,1), 0.7 & 0.18 (0.14) & 0.23 & 0.34 & 11.2 & 0 & 92 & 172\\18one LLM (Qwen2.5-7B) & (1,3), 0.0 & 0.27 (0.15) & 0.30 & 0.40 & 10.2 & 0 & 98 & 203\\19one LLM (Qwen2.5-7B) & (1,3), 0.7 & 0.16 (0.25) & 0.23 & 0.58 & 10.6 & 0 & 100 & 203\\20one LLM, five prompt variants & (0,1), 0.0 & 0.09 (0.08) & 0.09 & 0.21 & 10.5 & 0 & 95 & 149\\21one LLM, five prompt variants & (0,1), 0.7 & 0.10 (0.07) & 0.12 & 0.28 & 12.3 & 0 & 99 & 147\\22one LLM, five prompt variants & (1,1), 0.0 & 0.12 (0.04) & 0.13 & 0.27 & 11.4 & 0 & 84 & 147\\23one LLM, five prompt variants & (1,1), 0.7 & 0.26 (0.18) & 0.29 & 0.50 & 10.8 & 0 & 97 & 147\\24one LLM, five prompt variants & (1,3), 0.0 & 0.15 (0.08) & 0.16 & 0.37 & 10.4 & 0 & 97 & 144\\25one LLM, five prompt variants & (1,3), 0.7 & 0.10 (0.02) & 0.12 & 0.30 & 11.9 & 0 & 91 & 146\\26surrogate (rule, LLM interface) & (0,1), 0.0 & 0.09 (0.07) & 0.35 & 0.01 & 0.0 & 0 & 71 & 1\\27surrogate (rule, LLM interface) & (1,1), 0.0 & 1.00 (0.00) & 1.00 & 0.00 & 0.0 & 0 & 67 & 1\\28surrogate (rule, LLM interface) & (1,3), 0.0 & 0.69 (0.05) & 0.70 & 0.01 & 0.2 & 0 & 96 & 1\\29\bottomrule\end{tabular}30\begin{tablenotes}\footnotesize\item Means over seeds (s.d.\ in parentheses). Invalid: LLM replies rejected by the schema after one repair attempt (treated as hold); clipped: actions reduced by cash/holdings/margin limits. Source: \texttt{scripts/35\_llm\_lab.py}; audit logs in \texttt{results/llm\_audit}.\end{tablenotes}\end{table}31