% ============================================================================
% Appendix
% ============================================================================
\newpage
\begin{appendices}

\renewcommand{\thetable}{A\arabic{table}}
\renewcommand{\thefigure}{A\arabic{figure}}
\setcounter{table}{0}
\setcounter{figure}{0}

\section{Machine-Generated Feature Definitions}
\label{app:features}

Table~\ref{tab:app_features} lists the fields extracted by the language and
vision models under their respective strict JSON schemas, with the value domains as
specified in the extraction scripts (the enumerations are in French in the schema;
they are translated here). All fields use closed enumerations, integers, or
booleans; absent information is coded as \texttt{false}, \texttt{unknown}, or an
empty list. The mapping from fields to regressors is described in
Section~\ref{subsec:vision}.

\begin{table}[htbp]
  \centering
  \begin{threeparttable}
    \caption{Fields Extracted by the AI Models}
    \label{tab:app_features}
    \small
    \begin{tabular}{@{}p{3.1cm}p{4.3cm}p{6.2cm}@{}}
      \toprule
      Layer & Field & Domain \\
      \midrule
      \multicolumn{3}{@{}l}{\textit{Panel A: Text (LLM, \texttt{gpt-4o-mini}; 20 fields, 18 regressors)}} \\[3pt]
      Condition & \texttt{condition} & to renovate / original / good / renovated / like-new / new / unknown \\
      Turnkey; recently renovated & \texttt{turnkey}, \texttt{recently\_renovated} & boolean \\
      Renovated kitchen / bathroom & \texttt{renovated\_kitchen}, \texttt{renovated\_bathroom} & boolean \\
      Flooring (multi-valued) & \texttt{flooring} & hardwood / ceramic / floating / carpet / vinyl / concrete / other \\
      Countertop & \texttt{countertop} & granite / quartz / marble / laminate / wood / other / unknown \\
      Heating (multi-valued) & \texttt{heating} & electric / baseboards / heat pump / wood / gas / oil / radiant / geothermal / other \\
      View & \texttt{view} & water / mountain / city / none / unknown \\
      Amenities & \texttt{garage}, \texttt{pool}, \texttt{waterfront}, \texttt{finished\_basement}, \texttt{fireplace}, \texttt{central\_ac} & boolean \\
      New construction; income property & \texttt{new\_construction}, \texttt{income\_property} & boolean \\
      Large lot; quiet area & \texttt{large\_lot}, \texttt{quiet\_area} & boolean \\
      Highlights & \texttt{highlights} & 3--6 free-text keywords (not used in the regressions) \\[6pt]
      \multicolumn{3}{@{}l}{\textit{Panel B: Vision (\texttt{claude-haiku-4-5}, façade; 12 fields, 21 regressors)}} \\[3pt]
      Exterior shot & \texttt{is\_exterior} & boolean \\
      Building type (visual) & \texttt{building\_type\_visual} & detached / semi-detached / row / condo building / cottage / land / other \\
      Condition score & \texttt{condition\_score} & integer 1--10 \\
      Curb appeal & \texttt{curb\_appeal} & integer 1--10 \\
      Renovation level & \texttt{renovation\_level} & to renovate / dated / standard / renovated / high-end / unknown \\
      Exterior materials (multi-valued) & \texttt{exterior\_materials} & brick / stone / vinyl / wood / aluminum / stucco / roughcast / concrete / other \\
      Landscaping & \texttt{landscaping} & none / minimal / maintained / manicured \\
      Garage visible & \texttt{has\_garage\_visible} & boolean \\
      Storeys (visual) & \texttt{storeys\_visual} & 1 / 1.5 / 2 / 3+ / unknown \\
      Setting & \texttt{setting} & dense urban / suburban / rural / waterfront / wooded / unknown \\
      Photo quality & \texttt{photo\_quality} & integer 1--10 \\
      Luxury indicator & \texttt{curb\_luxury} & boolean \\
      \bottomrule
    \end{tabular}
    \begin{tablenotes}[flushleft]
      \footnotesize
      \item \textit{Notes:} Both extractions use schema-constrained decoding with
      closed enumerations, required fields, and \texttt{additionalProperties:false}.
      The text layer reads the free-text description (truncated at 4{,}000 characters)
      at temperature zero; the vision layer analyzes one façade photograph per listing
      at the API's default temperature. Model identifiers are the API aliases current in
      July 2026. Fields are mapped to numeric regressors (booleans to $0/1$, ordinal
      scales to integers, multi-valued fields and categories to indicators); the text
      condition scale is coded $0$--$5$ with ``unknown'' at~2, the visual renovation
      level $0$--$4$ with ``unknown'' at~2, and landscaping $0$--$3$.
    \end{tablenotes}
  \end{threeparttable}
\end{table}

\section{Sample Construction Flow}
\label{app:flow}

\begin{table}[htbp]
  \centering
  \begin{threeparttable}
    \caption{From Raw Records to Estimation Sample}
    \label{tab:app_flow}
    \begin{tabular}{@{}lr@{}}
      \toprule
      Step & Records \\
      \midrule
      Realtor.ca raw & 27{,}302 \\
      DuProprio raw & 7{,}693 \\
      Houses (single-family + townhouse) with price & 19{,}273 \\
      After outlier cleaning (price bounds, price-per-ft$^2$ trim) & 19{,}123 \\
      \quad of which Realtor.ca / DuProprio & 15{,}356 / 3{,}767 \\
      \quad of which with measured living area (Model~A) & 9{,}446 \\
      \quad of which with a detected exterior photograph & 18{,}314 \\
      \bottomrule
    \end{tabular}
    \begin{tablenotes}[flushleft]
      \footnotesize
      \item \textit{Notes:} See Section~\ref{sec:data}. ``Measured living area'' means
      reported area or area reconstructed from room-level dimensions. Both AI layers
      cover all 19{,}123 cleaned listings.
    \end{tablenotes}
  \end{threeparttable}
\end{table}

\section{Language Classification}
\label{app:language}

Each description is classified from the counts $e$ and $f$ of English and French
stop words it contains (articles, prepositions, pronouns, and auxiliaries such as
\textit{the, and, with, is} and \textit{le, la, les, et, avec, est}). A description is
``English'' if $e\ge1.5f$, ``French'' if $f\ge1.5e$, and ``mixed'' otherwise or when
neither count is positive. On Realtor.ca the mixed class is small (2.6\%) and
consists mostly of very short descriptions; on DuProprio it is large (51.7\%)
because the crawler returns the automated English translation followed by the
original French comments. Table~\ref{tab:rob_language_source} reports the gains from
the AI layers within each class, and Table~\ref{tab:rob_ai_coefs} the effect of
adding the language indicators to the hedonic regression. The language variable is
produced by \texttt{scripts/18\_revision\_language\_leakage\_spatial.py}.

\end{appendices}