spb/firc0001_uqo

1% Generated by code/make_tables.py from data/confidence\_dist.csv, results/ledger/code\_vs\_data\_checks.json. Do not edit by hand.2\begin{table}[H]3\centering\small4\caption{Self-reported extraction confidence of stored mentions, by run}\label{tab:confidence}5\begin{threeparttable}6\begin{tabular}{lrrrr}7\toprule8\multicolumn{1}{c}{\makecell{Self-reported\\confidence (rounded)}} & Mentions & \multicolumn{1}{c}{\makecell{Share\\(\%)}} & \multicolumn{1}{c}{\makecell{June run\\(backfilled)}} & \multicolumn{1}{c}{\makecell{October\\run}} \\9\midrule100.5 & 1{,}611 & 0.5 & 106 & 1{,}505 \\110.6 & 3{,}278 & 1.1 & 304 & 2{,}974 \\120.7 & 21{,}647 & 7.0 & 2{,}239 & 19{,}408 \\130.8 & 72{,}280 & 23.5 & 8{,}594 & 63{,}686 \\140.9 & 100{,}831 & 32.7 & 14{,}459 & 86{,}372 \\151.0 & 108{,}418 & 35.2 & 14{,}228 & 94{,}190 \\16\midrule17Total & 308{,}065 & 100.0 & 39{,}930 & 268{,}135 \\18\bottomrule19\end{tabular}20\begin{tablenotes}\footnotesize21\item Confidence is the value returned by the model in its JSON answer, rounded to one decimal; mentions below 0.45 are discarded before storage. A missing confidence is set to 0.6 by the post-processing code and passes the threshold: 3{,}278 mentions carry exactly 0.600 and cannot be separated from genuine 0.6 answers because the raw response is not stored. The distribution describes the model's reporting habit, not accuracy; calibration requires the human labels of gate (d) (Table~\ref{tab:gates}).22\end{tablenotes}23\end{threeparttable}24\end{table}25