% Generated by code/make_tables.py from results/grounding/grounding\_by\_group.csv, grounding\_element\_level.csv, grounding\_all\_structured.csv, fallback\_bounds.json. Do not edit by hand.
\begin{table}[H]
\centering\footnotesize
\setlength{\tabcolsep}{4pt}
\caption{Mechanical grounding of the structured fields of all stored mentions in the stored section text}\label{tab:grounding}
\begin{threeparttable}
\begin{adjustbox}{max width=\textwidth}
\begin{tabular}{lrrrrr}
\toprule
\multicolumn{6}{l}{\itshape Panel A: share of mentions whose field is found literally in the stored section text (\%, 95\% CI)} \\
Field & Mentions with the field & All forms & 10-K & 10-Q & 8-K \\
\midrule
Cited years & 252{,}469 & 99.8 [99.7, 99.8] & 99.6 & 99.8 & 99.8 \\
Partners & 81{,}320 & 99.4 [99.3, 99.5] & 99.3 & 99.1 & 99.6 \\
Amount (number) & 120{,}240 & 93.7 [93.6, 93.9] & 97.4 & 97.5 & 86.9 \\
Amount (number + currency token within 40 chars) & 120{,}240 & 80.5 [80.2, 80.7] & 86.8 & 82.3 & 74.9 \\
Locations & 118{,}783 & 78.3 [78.0, 78.6] & 85.7 & 82.1 & 67.8 \\
Technology tags & 36{,}740 & 63.6 [63.0, 64.2] & 63.8 & 64.8 & 58.2 \\
\midrule
\multicolumn{6}{l}{\itshape Panel B: element level (every listed year, partner, location or tag counted separately)} \\
Partners & 122{,}590 elements & 99.6 & & & \\
Locations & 181{,}898 elements & 84.2 & & & \\
Technology tags & 47{,}361 elements & 70.6 & & & \\
Cited years & 307{,}399 elements & 99.7 & & & \\
\midrule
\multicolumn{6}{l}{\itshape Panel C: mentions with every structured field grounded (\%), by form, run and confidence bin} \\
All mentions with at least one structured field & 276{,}168 & 88.2 & 92.2 & 92.4 & 76.6 \\
By run: June (backfilled) / October & 36{,}806 / 239{,}362 & 84.5 / 88.7 & & & \\
\quad confidence bin 0.45-0.6 & 696 & 90.7 & & & \\
\quad confidence bin 0.6-0.7 & 1{,}726 & 93.5 & & & \\
\quad confidence bin 0.7-0.8 & 13{,}313 & 94.2 & & & \\
\quad confidence bin 0.8-0.9 & 58{,}586 & 92.7 & & & \\
\quad confidence bin 0.9-1.0 & 201{,}847 & 86.4 & & & \\
\quad confidence bin 10-K & 69{,}065 & 92.2 & & & \\
\quad confidence bin 10-Q & 134{,}391 & 92.4 & & & \\
\quad confidence bin 8-K & 72{,}712 & 76.6 & & & \\
\bottomrule
\end{tabular}
\end{adjustbox}
\begin{tablenotes}\footnotesize
\item Population: all 308{,}065 stored mentions, joined to the text of their section as stored (legacy June text for the S\&P~500 10-Ks). A field is ``grounded'' when every element of the list (years, partners, locations, technologies, case-folded) or the amount (tested in units, thousands, millions and billions with one to three significant digits) occurs literally in that text; a currency token is \$, USD, dollar, million or billion within 40 characters of the number. Confidence intervals are percentile bootstraps with 1{,}000 replicates clustered by section. Grounding is a necessary condition for correctness, not a measure of it: a grounded year can belong to another initiative, and an ungrounded location can be a correct normalisation (``Texas'' for ``Austin''); precision and recall require the human labels of gate (d). Panel C counts mentions with at least one structured field. Upper bounds on the regex back-fill from the model's own paraphrase (\texttt{llm.py}): the stored years equal the regex output on the description for 98{,}438 of 252{,}469 mentions with years (39\%) and the technology list for 16{,}949 of 36{,}740 (46\%); 604 mentions cite a year absent from the section text.
\end{tablenotes}
\end{threeparttable}
\end{table}