| |
| |
| \begin{table*}[t] |
| \centering\small |
| \begin{tabular}{@{}l rrrr@{\hskip 14pt} rrrr@{}} |
| \toprule |
| & \multicolumn{4}{c}{10 structured fields} & \multicolumn{4}{c}{Cost block (4 fields)} \\ |
| \cmidrule(lr){2-5}\cmidrule(l){6-9} |
| System & Recall & Precision & F1 & False fill & Recall & Precision & F1 & False fill \\ |
| \midrule |
| Gemini & 65.9\,$\pm$\,0.5\% & 65.5\,$\pm$\,0.5\% & 0.66 & 30.2\,$\pm$\,0.6\% & \textbf{66.4\,$\pm$\,0.9\%} & 66.6\,$\pm$\,0.9\% & 0.67 & 17.4\,$\pm$\,0.7\% \\ |
| GPT-5.4-mini & \textbf{66.7\,$\pm$\,0.5\%} & 60.9\,$\pm$\,0.5\% & 0.64 & 41.8\,$\pm$\,0.7\% & 65.7\,$\pm$\,0.9\% & 59.6\,$\pm$\,0.9\% & 0.62 & 25.0\,$\pm$\,0.8\% \\ |
| Harvey & 58.8\,$\pm$\,0.5\% & 74.2\,$\pm$\,0.5\% & 0.66 & 16.2\,$\pm$\,0.5\% & 62.8\,$\pm$\,1.0\% & 69.4\,$\pm$\,1.0\% & 0.66 & 12.9\,$\pm$\,0.6\% \\ |
| Legora & 60.3\,$\pm$\,0.5\% & \textbf{82.6\,$\pm$\,0.5\%} & \textbf{0.70} & \textbf{8.7\,$\pm$\,0.4\%} & 61.9\,$\pm$\,1.0\% & \textbf{84.4\,$\pm$\,0.9\%} & \textbf{0.71} & \textbf{5.6\,$\pm$\,0.4\%} \\ |
| \bottomrule |
| \end{tabular} |
| \vskip 0.05in |
| \caption{Headline extraction metrics over the 19 release jurisdictions (8 core and 11 preview), computed over each system's successfully processed cases (metric definitions in \cref{sec:systems}). The left block covers the ten structured fields, the right block the four cost-block variables; percentages carry $\pm$1\,SE. Denominators, in row order (Gemini/GPT-5.4-mini/Harvey/Legora), are $n_{\text{filled}}$\,=\,10276/10276/8999/8132 and $n_{\text{empty}}$\,=\,5194/5184/4781/4438 over the ten structured fields, and $n_{\text{filled}}$\,=\,2896/2896/2491/2172 and $n_{\text{empty}}$\,=\,3292/3288/3021/2856 over the cost block; Harvey and Legora have smaller denominators because of their ingest gaps. F1 standard errors are below 0.01 and omitted. |
| The best value per column is marked in \textbf{bold} (lower is better for false fill).} |
| \label{tab:overall} |
| \end{table*} |
|
|
