| \begin{table}[t] | |
| \centering\small | |
| \caption{Sampling-strategy baselines. All three adapters are trained with identical QLoRA hyperparameters on the same underlying problem set; the only difference is which problems are drawn. Accuracy in \%; 95\% bootstrap CI in brackets; $p$-value from paired bootstrap against the quantized baseline.} | |
| \label{tab:baselines} | |
| \begin{tabular}{@{}lccc@{}} | |
| \toprule | |
| Sampling strategy & Acc (\%) & 95\% CI & $p$ \\ | |
| \midrule | |
| \textbf{Silver bullet} (ours) & 75.4 & [71.4, 79.2] & 0.001$^{**}$ \\ | |
| Failed only (no type balancing) & 74.0 & [70.0, 77.8] & 0.009$^{**}$ \\ | |
| Random (no diagnosis) & 73.6 & [69.6, 77.4] & 0.028$^{*}$ \\ | |
| \midrule | |
| \textit{Quantized baseline (no restoration)} & 69.4 & -- & -- \\ | |
| \bottomrule | |
| \end{tabular} | |
| \end{table} | |