1\begin{table}[htbp]\centering\footnotesize2\caption{Audit of the rule-based category labels (500-listing stratified sample)}\label{tab:cat_audit}3\begin{adjustbox}{max width=\textwidth}4\begin{tabular}{lrr}\toprule5Rule label & Listings with annotator consensus & Precision of the rule label \\\midrule6Maple \& honey & 24 & 0.50 \\7Fine grocery & 24 & 0.42 \\8Coffee \& tea & 23 & 0.35 \\9Chocolate & 24 & 0.62 \\10Bakery & 25 & 0.40 \\11Meat \& fish & 22 & 0.32 \\12Cheese \& dairy & 24 & 0.58 \\13Beverages & 24 & 0.17 \\14Alcohol & 25 & 0.28 \\15Beauty \& care & 24 & 0.62 \\16Home \& decor & 22 & 0.36 \\17Fashion & 25 & 0.72 \\18Jewellery & 25 & 0.80 \\19Art \& craft & 25 & 0.52 \\20Children & 21 & 0.62 \\21Pets & 24 & 0.29 \\22Outdoor \& sport & 19 & 0.32 \\23Garden & 25 & 0.84 \\24Industrial & 24 & 0.42 \\25Other & 25 & 0.00 \\26\textbf{All} & 474 & \textbf{0.46} \\27\bottomrule\end{tabular}28\end{adjustbox}29\begin{minipage}{0.95\textwidth}\vspace{2pt}\footnotesize\textit{Notes:} 25 random listings per category annotated independently by two LLMs (claude-opus-5-5, gpt-5.5) with the platform's taxonomy; agreement rule vs A 0.45, rule vs B 0.45, A vs B 0.95; consensus exists for 95\% of listings. Precision = share of consensus labels equal to the rule label. Disagreements listed in the replication package.\end{minipage}30\end{table}31