% Author: Simon-Pierre Boucher — simon-pierre.boucher@uqo.ca % generated by scripts/20_evaluate.py — do not edit \begin{tabular}{lrrrrrr} \toprule Arm & {Cost (USD)} & {Latency (s)} & {Tokens} & {Total cost (USD)} & {Mean $|e|$} & {N} \\ \midrule A (hedonic) & \num{0.0000} & \num{0.0} & \num{0} & \num{0.00} & \num{0.197} & \num{72132} \\ B (hierarchical) & \num{0.0000} & \num{0.0} & \num{0} & \num{0.00} & \num{0.190} & \num{72132} \\ C (catboost) & \num{0.0000} & \num{0.0} & \num{0} & \num{0.00} & \num{0.173} & \num{72132} \\ C (gbm) & \num{0.0000} & \num{0.0} & \num{0} & \num{0.00} & \num{0.183} & \num{72132} \\ C (xgboost) & \num{0.0000} & \num{0.0} & \num{0} & \num{0.00} & \num{0.181} & \num{72132} \\ D (avm) & \num{0.0000} & \num{0.0} & \num{0} & \num{0.00} & \num{0.180} & \num{72132} \\ \midrule E (single llm) & \num{0.0206} & \num{19.6} & \num{6381} & \num{0.83} & \num{0.182} & \num{40} \\ F (independent) & \num{0.2269} & \num{84.7} & \num{71977} & \num{9.08} & \num{0.179} & \num{40} \\ G (debate r1) & \num{0.8773} & \num{252.2} & \num{294497} & \num{35.09} & \num{0.178} & \num{40} \\ G (debate r2) & \num{1.5012} & \num{402.3} & \num{514936} & \num{60.05} & \num{0.176} & \num{40} \\ G (debate r4) & \num{2.2712} & \num{575.3} & \num{776036} & \num{45.42} & \num{0.163} & \num{20} \\ G (fullview) & \num{1.5952} & \num{411.5} & \num{545943} & \num{63.81} & \num{0.175} & \num{40} \\ H (debate skeptic) & \num{1.3740} & \num{383.2} & \num{464506} & \num{54.96} & \num{0.175} & \num{40} \\ I (placebo) & \num{1.3519} & \num{317.8} & \num{476454} & \num{54.08} & \num{0.184} & \num{40} \\ J (cost matched) & \num{0.2523} & \num{107.9} & \num{89422} & \num{10.09} & \num{0.179} & \num{40} \\ \bottomrule \end{tabular}