paper/sections/appendix.tex

% ============================================================
%  APPENDIX
% ============================================================
\appendix

% Copy-paste safety: these definitions are no-ops when main.tex already defines them.
\providecommand{\ms}[2]{\ensuremath{#1{\mkern1mu}_{\scriptscriptstyle \pm #2}}}


% ────────────────────────────────────────────────────────────
%  TABLE OF CONTENTS (appendix only)
% ────────────────────────────────────────────────────────────
\section*{Appendix Contents}
\addcontentsline{toc}{section}{Appendix Contents}


\begin{center}
\begin{tabular}{@{}p{0.82\textwidth}r@{}}
\textbf{A\quad Evaluation Contract Specifications} & \pageref{sec:app_contract} \\
\quad A.1\enspace Matching Rule Definitions        & \pageref{sec:app_contract_matching} \\
\quad A.2\enspace Task-Form Contract Parameters    & \pageref{sec:app_contract_params} \\
\quad A.3\enspace Evaluation Scope Definitions     & \pageref{sec:app_contract_scope} \\[4pt]
\textbf{B\quad Controlled Check Details}           & \pageref{sec:app_checks} \\
\quad B.1\enspace Fixed-Output Check: Full Sweep   & \pageref{sec:app_checks_output} \\
\quad B.2\enspace Fixed-Feature Check: Selection Summary & \pageref{sec:app_checks_feature} \\
\quad B.3\enspace Selection Regret Under Matching Rules & \pageref{sec:app_checks_regret} \\
\quad B.4\enspace Additional Value Tables & \pageref{sec:app_checks_values} \\[4pt]
\textbf{C\quad Comparator Eligibility Notes}       & \pageref{sec:comparator_audit} \\[4pt]
\textbf{D\quad Seeded Audits}                      & \pageref{sec:app_seeded_audits} \\
\quad D.1\enspace Seed Robustness Summary          & \pageref{sec:app_seed_robustness} \\[4pt]
\textbf{E\quad Lightweight Head and Adaptation Details} & \pageref{sec:app_heads} \\[4pt]
\textbf{F\quad Limitations}                        & \pageref{sec:limitations} \\[4pt]
\textbf{G\quad Reproducibility and Evaluation Artifacts} & \pageref{sec:repro_compute_impact} \\
\end{tabular}
\end{center}

\noindent\textit{Retention rule.}
Appendix tables are retained when they add contract parameters, controlled-check arithmetic,
task-specific non-main metrics, seed summaries, eligibility checks, or protocol details.
Full task matrices and reference-summary tables that repeat the main result tables are not repeated here.

\clearpage

% ============================================================
%  A  EVALUATION CONTRACT SPECIFICATIONS
% ============================================================
\section{Evaluation Contract Specifications}
\label{sec:app_contract}

% ────────────────────────────────────────────────────────────
\subsection{Matching Rule Definitions}
\label{sec:app_contract_matching}

The three matching rules used across occupancy task forms are defined as follows.

\noindent\textbf{Exact matching.}
A predicted unit-time pair $(i,t) \in \widehat{P}_\tau$ is counted as a true positive if and only if the same pair appears in the observed fire set $P = \{(i,t): y_{i,t}=1\}$.
This is the strictest rule and yields the lowest $F_1$ for any fixed output.

\noindent\textbf{Tolerated matching.}
A predicted pair $(i,t)$ is counted as correct if there exists an observed pair $(i',t') \in P$ such that $\|i - i'\|_\infty \le k$ and $|t - t'| \le \Delta t$, where $k$ is the spatial tolerance in grid cells and $\Delta t$ is the temporal tolerance in forecast steps.
Both parameters are fixed as part of the evaluation contract $\mathcal{C}$ before scoring.

\noindent\textbf{Union matching.}
A predicted pair is counted as a true positive if it satisfies either exact or tolerated matching.
The resulting union-$F_1$ provides an upper bound on decision performance under the chosen tolerance.

\noindent\textbf{Fixed parameter values.}
For occupancy, the spatial tolerance is $k=8$ grid cells.
The temporal tolerance is $\Delta t=3$ forecast steps for union matching and $\Delta t=0$ for spatial-only tolerance.
The threshold $\tau$ is selected on validation strict-$F_1$ before test scoring.
For fire spread, the spatial tolerance is $k=4$ grid cells, $\Delta t=0$, and the threshold is selected on validation spatial $F_1$.

\noindent Table~\ref{tab:app_matching_rule_params} records the fixed matching-rule parameters.

\begin{table}[h]
\centering
\small
\setlength{\tabcolsep}{10pt}
\renewcommand{\arraystretch}{1.2}
\caption{Matching-rule values used in the evaluation contracts.}
\label{tab:app_matching_rule_params}
\begin{tabular}{lll}
\toprule
\textbf{Parameter} & \textbf{Occupancy} & \textbf{Fire spread} \\
\midrule
\(k\) & 8 cells & 4 cells \\
\(\Delta t\) & 3 for union; 0 spatial-only & 0 \\
\(\tau\) & val. strict \(F_1\) & val. spatial \(F_1\) \\
\bottomrule
\end{tabular}
\end{table}

% ────────────────────────────────────────────────────────────
\subsection{Task-Form Contract Parameters}
\label{sec:app_contract_params}

Table~\ref{tab:app_contract_params_full} lists fixed scoring values not shown in the main contract map.

\begin{table}[h]
\centering
\scriptsize
\setlength{\tabcolsep}{3.5pt}
\renewcommand{\arraystretch}{1.2}
\caption{Fixed scoring values used by each task-form contract.}
\label{tab:app_contract_params_full}
\begin{adjustbox}{max width=\textwidth}
\begin{tabular}{llll}
\toprule
\textbf{\(\mathcal{T}\)} & \textbf{Scoring} & \textbf{Validation} & \textbf{\(\Omega\)} \\
\midrule
Occupancy & \(k=8,\Delta t=3\); exact/tol./union \(F_1\) & val. strict \(F_1\) & global; top-5/10/20\% fire-prone \\
Fire spread & \(k=4,\Delta t=0\); exact/spatial \(F_1\), AP & val. spatial \(F_1\) & spread-region cells \\
Final burned area & log-RMSE, log-MAE, Spearman \(\rho\) & val. log-RMSE & test events \\
Analog retrieval & nDCG@10; retrieved-event log error & val. nDCG@10 & test events \\
Smoke PM\(_{2.5}\) & RMSE, MAE, Pearson \(r\); exceedance 35 & val. RMSE & test stations \\
Extreme heat & RMSE-C, MAE-C, exceedance \(F_1\) & val. threshold 27/30/33\(^{\circ}\)C & heat-region stations \\
\bottomrule
\end{tabular}
\end{adjustbox}
\end{table}

% ────────────────────────────────────────────────────────────
\subsection{Evaluation Scope Definitions}
\label{sec:app_contract_scope}

\noindent\textbf{Global scope.}
Evaluation covers all spatial units in the domain, including fire-inactive regions.
This scope can mask model differences on fire-relevant locations because inactive cells inflate true-negative counts.

\noindent\textbf{Fire-prone scope.}
Evaluation is restricted to grid cells in the top-$k$\% of historical fire activity.
We report results for top-5\%, top-10\%, and top-20\% cutoffs.
The cutoff thresholds are derived from the training period and held fixed at test time.

\noindent\textbf{Spread region scope.}
For fire spread tasks, evaluation is restricted to the predicted and observed burned raster patches.
Only cells within the union of $\widehat{B}$ and $B$ contribute to metric computation.

\noindent\textbf{Fixed scope sizes.}
The global scope contains 8,085,000 test cells.
The fire-prone top-5\%, top-10\%, and top-20\% scopes contain 404,280, 808,560, and 1,617,000 test cells, respectively.
The spread-region scope is event-specific and uses the union of $\widehat{B}$ and $B$.

\begin{table}[h]
\centering
\small
\setlength{\tabcolsep}{8pt}
\renewcommand{\arraystretch}{1.2}
\caption{Scope values used in the evaluation contracts.}
\label{tab:app_scope_params}
\begin{tabular}{lcc}
\toprule
\textbf{\(\Omega\)} & \textbf{Definition} & \textbf{Units} \\
\midrule
Global & full domain & 8,085,000 test cells \\
Fire-prone top-5\% & top 5\% by training-period fire frequency & 404,280 test cells \\
Fire-prone top-10\% & top 10\% by training-period fire frequency & 808,560 test cells \\
Fire-prone top-20\% & top 20\% by training-period fire frequency & 1,617,000 test cells \\
Spread region & union of \(\widehat{B}\) and \(B\) & event-specific cells \\
\bottomrule
\end{tabular}
\end{table}

\clearpage

% ============================================================
%  B  CONTROLLED CHECK DETAILS
% ============================================================
\section{Controlled Check Details}
\label{sec:app_checks}


\begin{figure}[t]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_fireprone_contract_progression_compact.pdf}
    \caption{
\textbf{Matching-rule sensitivity in fire-prone occupancy (RQ1).}
Each row holds the score field \(S\), label field \(Y\), threshold, and \(\Omega\) fixed, and changes only \(\Lambda\).
Legend: \textcolor[HTML]{17375E}{$\blacksquare$} strict \(F_1\),
\textcolor[HTML]{4F8DCC}{$\blacksquare$} added \(F_1\) from spatial tolerance,
\textcolor[HTML]{BFD7F0}{$\blacksquare$} added \(F_1\) from union matching,
red outline \ourfm, and dashed line original weather FMs vs.\ added baselines.
The horizontal axis is \(F_1\) in percent.
}
    \label{fig:fireprone_contract_progression}
\end{figure}

% ────────────────────────────────────────────────────────────
\subsection{Fixed-Output Check: Full Sweep}
\label{sec:app_checks_output}

The fixed-output check holds the score field $S$ and label field $Y$ fixed and varies only $\Lambda$.
Table~\ref{tab:fireprone_contract_progression} reports the full global and fire-prone sweep for all retained backbones.
The same table is the numeric counterpart to Figure~\ref{fig:fireprone_contract_progression}.

\begin{table*}[t]
    \centering
    \scriptsize
    \setlength{\tabcolsep}{4pt}
    \caption{Occupancy \(F_1\) scores across global and fire-prone scopes. Global uses the full validation/test domain; top-\(k\) rows use train-defined fire-prone masks from historical fire frequency. Values are percentages from the same validation-selected strict threshold. Tolerance is spatial-only; union adds temporal and spatial matching. \(\Delta\) is union minus strict. Cells report five-seed mean with std in small type.}
    \label{tab:fireprone_contract_progression}
    \begin{tabular}{@{}llcccc@{}}
        \toprule
        Backbone & \(\Omega\) & Strict \(F_1\uparrow\) & Tol.\ \(F_1\uparrow\) & Union \(F_1\uparrow\) & \(\Delta\) \(\uparrow\) \\
        \midrule
        \ourfm & global & \ms{0.4546}{0.1412} & \ms{29.7484}{1.2868} & \ms{59.0656}{2.7372} & \ms{58.6109}{2.6945} \\
         & top 5\% & \ms{3.5604}{0.8809} & \ms{39.2617}{1.4011} & \ms{72.8280}{2.5784} & \ms{69.2676}{1.9960} \\
         & top 10\% & \ms{3.5575}{0.8799} & \ms{39.1665}{1.3906} & \ms{72.5204}{2.5670} & \ms{68.9629}{1.9888} \\
         & top 20\% & \ms{3.5300}{0.8700} & \ms{38.2849}{1.2952} & \ms{69.7228}{2.4664} & \ms{66.1928}{1.9273} \\
        \addlinespace[1pt]
        Prithvi-WxC & global & \ms{0.0552}{0.0039} & \ms{7.1649}{0.6557} & \ms{20.1853}{1.8299} & \ms{20.1301}{1.8297} \\
         & top 5\% & \ms{1.4119}{1.1635} & \ms{19.2636}{4.5019} & \ms{42.5793}{4.5495} & \ms{41.1674}{3.4846} \\
         & top 10\% & \ms{1.2376}{1.3201} & \ms{14.8780}{8.4429} & \ms{32.6913}{13.2085} & \ms{31.4536}{11.9053} \\
         & top 20\% & \ms{1.1520}{1.3770} & \ms{13.1512}{9.4556} & \ms{28.1319}{15.2866} & \ms{26.9800}{13.9224} \\
        \addlinespace[1pt]
        Aurora & global & \ms{0.0656}{0.0094} & \ms{8.5009}{1.9594} & \ms{23.1037}{4.9418} & \ms{23.0382}{4.9325} \\
         & top 5\% & \ms{0.9859}{0.9299} & \ms{15.1337}{6.0821} & \ms{35.4834}{11.0192} & \ms{34.4975}{10.3728} \\
         & top 10\% & \ms{0.7790}{1.0453} & \ms{12.7381}{6.5558} & \ms{30.5305}{10.8842} & \ms{29.7515}{9.8656} \\
         & top 20\% & \ms{0.6655}{1.1043} & \ms{10.5304}{7.4309} & \ms{24.9444}{12.5844} & \ms{24.2790}{11.4943} \\
        \addlinespace[1pt]
        ClimaX & global & \ms{0.3480}{0.0754} & \ms{29.7535}{3.6073} & \ms{60.1506}{7.5865} & \ms{59.8026}{7.5454} \\
         & top 5\% & \ms{1.2937}{0.1086} & \ms{34.5791}{2.3772} & \ms{69.2186}{5.7215} & \ms{67.9249}{5.7263} \\
         & top 10\% & \ms{1.2522}{0.1602} & \ms{34.3341}{2.2852} & \ms{68.5713}{5.5377} & \ms{67.3191}{5.5538} \\
         & top 20\% & \ms{1.0287}{0.2686} & \ms{30.2140}{4.2857} & \ms{60.0650}{7.5674} & \ms{59.0363}{7.5891} \\
        \addlinespace[1pt]
        StormCast & global & \ms{0.0626}{0.0119} & \ms{8.1951}{2.1895} & \ms{22.3817}{5.4294} & \ms{22.3191}{5.4178} \\
         & top 5\% & \ms{0.9573}{0.8011} & \ms{15.3219}{5.5337} & \ms{36.1857}{9.7331} & \ms{35.2284}{9.1816} \\
         & top 10\% & \ms{0.7284}{0.9280} & \ms{12.6669}{6.3290} & \ms{30.4748}{10.6527} & \ms{29.7464}{9.7494} \\
         & top 20\% & \ms{0.5795}{0.9104} & \ms{10.4157}{7.3437} & \ms{24.6598}{12.3973} & \ms{24.0803}{11.4988} \\
        \addlinespace[1pt]
        DLWP & global & \ms{0.1693}{0.0419} & \ms{14.9148}{3.2446} & \ms{28.1901}{6.9658} & \ms{28.0208}{6.9257} \\
         & top 5\% & \ms{1.8054}{0.4835} & \ms{31.7231}{3.2923} & \ms{55.4596}{5.2920} & \ms{53.6542}{5.4752} \\
         & top 10\% & \ms{1.6110}{0.5999} & \ms{27.6581}{5.9216} & \ms{47.1269}{8.0111} & \ms{45.5158}{7.7927} \\
         & top 20\% & \ms{1.5248}{0.8987} & \ms{20.9403}{4.7971} & \ms{34.9301}{7.8471} & \ms{33.4054}{7.8760} \\
        \addlinespace[1pt]
        FCN & global & \ms{0.2829}{0.0839} & \ms{19.5061}{3.3412} & \ms{40.0604}{9.3701} & \ms{39.7775}{9.3423} \\
         & top 5\% & \ms{1.6231}{0.5064} & \ms{29.3769}{2.7626} & \ms{54.3033}{7.4089} & \ms{52.6801}{7.4389} \\
         & top 10\% & \ms{1.1777}{0.5118} & \ms{22.4217}{3.9803} & \ms{43.4510}{9.2513} & \ms{42.2734}{9.0251} \\
         & top 20\% & \ms{0.9962}{0.4315} & \ms{16.9792}{3.9371} & \ms{34.0859}{8.2616} & \ms{33.0897}{7.9275} \\
        \addlinespace[1pt]
        FengWu & global & \ms{0.2613}{0.0757} & \ms{12.0050}{6.0239} & \ms{24.1022}{13.6293} & \ms{23.8410}{13.5736} \\
         & top 5\% & \ms{1.5695}{0.3592} & \ms{16.2763}{3.7024} & \ms{30.1055}{5.0103} & \ms{28.5360}{4.7696} \\
         & top 10\% & \ms{1.2427}{0.5333} & \ms{12.9503}{5.6052} & \ms{24.1854}{8.6854} & \ms{22.9427}{8.1863} \\
         & top 20\% & \ms{1.1192}{0.5023} & \ms{11.9508}{5.0745} & \ms{22.7860}{7.9115} & \ms{21.6668}{7.4438} \\
        \addlinespace[1pt]
        FuXi & global & \ms{0.3774}{0.1212} & \ms{21.0323}{4.8211} & \ms{37.2888}{9.4470} & \ms{36.9114}{9.4327} \\
         & top 5\% & \ms{2.0307}{0.6800} & \ms{31.8944}{4.7331} & \ms{53.9308}{8.3822} & \ms{51.9001}{8.6878} \\
         & top 10\% & \ms{1.6542}{0.7316} & \ms{24.0128}{5.7784} & \ms{40.2140}{9.9307} & \ms{38.5597}{9.7744} \\
         & top 20\% & \ms{1.3646}{0.6773} & \ms{21.9548}{5.8601} & \ms{36.7314}{10.0289} & \ms{35.3668}{9.9223} \\
        \addlinespace[1pt]
        Pangu-Weather & global & \ms{0.2755}{0.1089} & \ms{17.0909}{4.0477} & \ms{35.6386}{9.0327} & \ms{35.3630}{9.0774} \\
         & top 5\% & \ms{1.3656}{0.3064} & \ms{22.2222}{6.8613} & \ms{43.4234}{13.2383} & \ms{42.0578}{13.0599} \\
         & top 10\% & \ms{1.0931}{0.3535} & \ms{18.9337}{5.9329} & \ms{38.5325}{11.7221} & \ms{37.4394}{11.5261} \\
         & top 20\% & \ms{0.8844}{0.3601} & \ms{17.0172}{5.4859} & \ms{34.5688}{10.2932} & \ms{33.6844}{10.1334} \\
        \addlinespace[1pt]
        AlphaEarth & global & \ms{2.0606}{0.4404} & \ms{29.4476}{6.0064} & \ms{37.4286}{9.9458} & \ms{35.3679}{10.0271} \\
         & top 5\% & \ms{6.9133}{0.8450} & \ms{42.8790}{4.6087} & \ms{51.7449}{8.7321} & \ms{44.8315}{9.0763} \\
         & top 10\% & \ms{6.6366}{0.9901} & \ms{41.8981}{5.9454} & \ms{50.5712}{10.0057} & \ms{43.9346}{9.9156} \\
         & top 20\% & \ms{6.1908}{1.1330} & \ms{38.8325}{7.4966} & \ms{46.3833}{12.1697} & \ms{40.1925}{11.6788} \\
        \bottomrule
    \end{tabular}
\end{table*}

% ────────────────────────────────────────────────────────────
\subsection{Fixed-Feature Check: Selection Summary}
\label{sec:app_checks_feature}

The paper appendix keeps the fixed-feature result at the selection-summary level.
The full per-head rows are retained in the supplementary CSV files and are not repeated as a manuscript table because many degenerate heads produce identical zero decision scores.
The supplementary selection rows report the decision-score loss after changing only the head-selection metric.

% ────────────────────────────────────────────────────────────
\subsection{Selection Regret Under Matching Rules}
\label{sec:app_checks_regret}

The fixed-feature check trains the same head family $\mathcal{H}$ on a fixed feature source and changes only the selection metric.
Table~\ref{tab:appendix_selection_regret_tolerance} reports the same selection comparison under exact, tolerated, and union matching.
Here, \(h_R\) is selected by PR-AUC and \(h_D\) is selected by the decision metric.
The reported regret is \(D(h_D)-D(h_R)\).
Exact zero entries mean the two selectors give the same decision score for all five seeds.

\begin{table*}[!t]
    \centering
    \scriptsize
    \setlength{\tabcolsep}{4pt}
    \caption{Selection-regret values under exact, tolerated, and union matching. Values are percentage-point regret from selecting \(h_R\) by PR-AUC instead of \(h_D\) by the decision metric. Rows report mean with small std over five seeds; \(0.0000\) denotes exact zero regret.}
    \label{tab:appendix_selection_regret_tolerance}
    \begin{adjustbox}{max width=\textwidth}
    \begin{tabular}{llccc}
        \toprule
        \textbf{Feature} & \textbf{\(\Omega\)} & \textbf{Exact regret} & \textbf{Tolerated regret} & \textbf{Union regret} \\
        \midrule
        \ourfm & global & 0.0000 & \ms{8.7830}{9.6705} & \ms{8.7830}{9.6705} \\
        \ourfm & fire-prone & 0.0000 & \ms{3.4027}{3.2045} & \ms{3.4027}{3.2045} \\
        Prithvi-WxC & global & 0.0000 & 0.0000 & 0.0000 \\
        Prithvi-WxC & fire-prone & 0.0000 & 0.0000 & 0.0000 \\
        Aurora & global & \ms{0.0200}{0.0267} & \ms{9.8520}{12.9878} & \ms{9.8520}{12.9878} \\
        Aurora & fire-prone & \ms{0.8203}{1.8341} & \ms{14.3919}{32.1219} & \ms{14.3919}{32.1219} \\
        ClimaX & global & \ms{0.0003}{0.0004} & \ms{0.1296}{0.1775} & \ms{0.1296}{0.1775} \\
        ClimaX & fire-prone & 0.0000 & 0.0000 & 0.0000 \\
        StormCast & global & 0.0000 & 0.0000 & 0.0000 \\
        StormCast & fire-prone & 0.0000 & 0.0000 & 0.0000 \\
        DLWP & global & 0.0000 & 0.0000 & 0.0000 \\
        DLWP & fire-prone & \ms{0.0770}{0.1100} & \ms{4.3266}{4.3323} & \ms{4.3266}{4.3323} \\
        FCN & global & 0.0000 & 0.0000 & 0.0000 \\
        FCN & fire-prone & \ms{0.0006}{0.0013} & \ms{1.1680}{1.9872} & \ms{1.1680}{1.9872} \\
        FengWu & global & 0.0000 & 0.0000 & 0.0000 \\
        FengWu & fire-prone & \ms{0.0691}{0.1191} & \ms{0.5222}{0.6239} & \ms{0.5222}{0.6239} \\
        FuXi & global & 0.0000 & 0.0000 & 0.0000 \\
        FuXi & fire-prone & 0.0000 & \ms{0.1084}{0.1729} & \ms{0.1084}{0.1729} \\
        Pangu-Weather & global & 0.0000 & 0.0000 & 0.0000 \\
        Pangu-Weather & fire-prone & \ms{0.0728}{0.1179} & \ms{0.1849}{0.3263} & \ms{0.1849}{0.3263} \\
        AlphaEarth & global & 0.0000 & \ms{17.2217}{8.8492} & \ms{17.2217}{8.8492} \\
        AlphaEarth & fire-prone & 0.0000 & \ms{3.8804}{5.9483} & \ms{3.8804}{5.9483} \\
        \bottomrule
    \end{tabular}
    \end{adjustbox}
\end{table*}


% ────────────────────────────────────────────────────────────
\subsection{Additional Value Tables}
\label{sec:app_checks_values}

Table~\ref{tab:app_occupancy_ppr_scope}
reports the predicted-positive rate behind the occupancy \(F_1\) sweep.

\begin{table*}[t]
\centering
\small
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.18}
\caption{For fixed occupancy \(\mathcal{T}\), this table reports predicted-positive rate.
Values are percentages under the same validation-selected strict threshold.
Scopes \(\Omega\) are fixed before test scoring; cells report five-seed mean with std in small type.}
\label{tab:app_occupancy_ppr_scope}
\begin{tabular}{lcccc}
\toprule
\textbf{Backbone} & \textbf{\(\Omega=\)global} & \textbf{\(\Omega=\)top 5\%} & \textbf{\(\Omega=\)top 10\%} & \textbf{\(\Omega=\)top 20\%} \\
\midrule
\ourfm & \ms{1.6808}{0.3684} & \ms{3.0619}{1.0925} & \ms{1.5310}{0.5463} & \ms{0.7655}{0.2732} \\
Prithvi-WxC & \ms{61.9711}{30.9101} & \ms{57.4117}{47.8987} & \ms{58.4565}{51.0897} & \ms{58.9788}{52.6991} \\
Aurora & \ms{55.5849}{19.7524} & \ms{57.2238}{35.3400} & \ms{68.7942}{37.6958} & \ms{67.2891}{38.3991} \\
ClimaX & \ms{5.6763}{3.9261} & \ms{24.0091}{9.2816} & \ms{11.8450}{4.5067} & \ms{5.7442}{4.1341} \\
StormCast & \ms{60.6507}{17.4895} & \ms{57.6017}{35.2921} & \ms{68.0766}{37.3899} & \ms{67.8397}{39.2410} \\
DLWP & \ms{4.3221}{1.5619} & \ms{9.4001}{5.0807} & \ms{4.9700}{3.6849} & \ms{1.9198}{1.4678} \\
FCN & \ms{1.5202}{1.3446} & \ms{4.7856}{2.9409} & \ms{2.7257}{1.6353} & \ms{0.8368}{0.2358} \\
FengWu & \ms{0.4277}{0.4830} & \ms{0.6004}{0.3041} & \ms{0.2609}{0.1935} & \ms{0.1501}{0.1206} \\
FuXi & \ms{0.4505}{0.2773} & \ms{2.9315}{2.6392} & \ms{0.5197}{0.6074} & \ms{0.3621}{0.4346} \\
Pangu-Weather & \ms{1.0801}{1.1308} & \ms{2.0549}{2.1893} & \ms{1.4029}{1.4739} & \ms{1.0103}{1.1084} \\
AlphaEarth & \ms{0.0691}{0.0499} & \ms{0.2826}{0.1497} & \ms{0.1524}{0.0770} & \ms{0.0656}{0.0414} \\
\bottomrule
\end{tabular}
\end{table*}

Tables~\ref{tab:app_spread_ap_by_scope}--\ref{tab:app_heat_event_pr}
report additional values that are not repeated in the main tables.
Each table fixes the task \(\mathcal{T}\) and reports either a different \(\Omega\), metric, or event subset.

\begin{table*}[t]
\centering
\scriptsize
\setlength{\tabcolsep}{3pt}
\caption{For fixed spread \(\mathcal{T}\) and strict \(\Lambda\), this table reports AP under three \(\Omega\) scopes: full test, top-5\% train-fire area, and top-10\% train-fire area. Values are percentages; cells report mean with small std.}
\label{tab:app_spread_ap_by_scope}
\begin{tabular}{lccc}
\toprule
Backbone & full \(\Omega\) AP & top-5\% \(\Omega\) AP & top-10\% \(\Omega\) AP \\
\midrule
\ourfm & \ms{30.0197}{1.5651} & \ms{40.7452}{2.0542} & \ms{37.4096}{1.8731} \\
Prithvi-WxC & \ms{4.8319}{0.1731} & \ms{12.6086}{0.4468} & \ms{8.7051}{0.1889} \\
Aurora & \ms{17.7723}{0.4293} & \ms{30.3106}{0.9404} & \ms{26.4732}{0.6932} \\
ClimaX & \ms{11.1726}{0.2337} & \ms{25.7871}{1.2896} & \ms{19.9977}{1.2217} \\
StormCast & \ms{8.1147}{1.1569} & \ms{18.5461}{1.1727} & \ms{14.1286}{1.2956} \\
DLWP & \ms{9.2142}{2.6587} & \ms{19.3346}{2.3922} & \ms{14.9788}{2.6696} \\
FCN & \ms{6.6774}{1.3001} & \ms{16.7396}{3.2955} & \ms{11.9308}{2.3881} \\
FengWu & \ms{11.0046}{2.7092} & \ms{21.1506}{1.2163} & \ms{17.0113}{1.5778} \\
FuXi & \ms{13.5507}{0.3840} & \ms{22.5434}{0.4100} & \ms{19.1964}{0.3943} \\
Pangu-Weather & \ms{10.6250}{1.4643} & \ms{19.8294}{1.3044} & \ms{15.8013}{1.1602} \\
AlphaEarth & \ms{12.2847}{1.3562} & \ms{22.8692}{0.4915} & \ms{18.2992}{1.2110} \\
\bottomrule
\end{tabular}
\end{table*}

\begin{table*}[t]
\centering
\scriptsize
\setlength{\tabcolsep}{3pt}
\caption{For fixed final-area \(\mathcal{T}\) and \(\Omega\), this table reports median log error and acre-scale errors in addition to the main log-RMSE/log-MAE/Spearman metrics. Cells report mean with small std.}
\label{tab:app_burned_area_median_acre}
\begin{tabular}{lccc}
\toprule
Backbone & log median AE & acre median AE & acre MAPE \\
\midrule
\ourfm & \ms{1.0235}{0.0982} & \ms{4504.0692}{459.0483} & \ms{1.4525}{0.0254} \\
Prithvi-WxC & \ms{1.2184}{0.2107} & \ms{5375.8770}{788.7906} & \ms{1.9517}{0.2875} \\
Aurora & \ms{1.4547}{0.0301} & \ms{9904.9483}{457.4260} & \ms{6.8728}{3.0026} \\
ClimaX & \ms{1.6841}{0.1818} & \ms{18130.4820}{3248.3873} & \ms{8.2373}{2.8540} \\
StormCast & \ms{1.4522}{0.1519} & \ms{11155.7881}{2020.8656} & \ms{4.6142}{1.1500} \\
DLWP & \ms{1.0952}{0.1306} & \ms{4406.9315}{303.0944} & \ms{1.7357}{0.3625} \\
FCN & \ms{1.1688}{0.1139} & \ms{5166.9993}{213.0333} & \ms{2.0800}{0.4004} \\
FengWu & \ms{1.1589}{0.1772} & \ms{5137.2822}{628.7543} & \ms{2.0944}{0.4545} \\
FuXi & \ms{1.1855}{0.0612} & \ms{5697.7117}{796.8785} & \ms{2.4411}{0.5567} \\
Pangu-Weather & \ms{1.1221}{0.1470} & \ms{5092.3621}{483.8243} & \ms{1.9571}{0.3113} \\
AlphaEarth & \ms{1.7459}{0.6057} & \ms{15110.7573}{7106.3417} & \ms{9.7398}{2.7425} \\
\bottomrule
\end{tabular}
\end{table*}

\begin{table*}[t]
\centering
\scriptsize
\setlength{\tabcolsep}{3pt}
\caption{For fixed retrieval \(\mathcal{T}\) and \(\Omega\), this table reports nDCG@5, best log gap, and rank \(\rho\) in addition to the main nDCG@10/log-error metrics. Cells report mean with small std.}
\label{tab:app_analog_rank_depth}
\begin{tabular}{lccc}
\toprule
Backbone & nDCG@5 & best log gap & rank $\rho$ \\
\midrule
\ourfm & \ms{0.5175}{0.0445} & \ms{0.1868}{0.0285} & \ms{0.6019}{0.1460} \\
Prithvi-WxC & \ms{0.3591}{0.0107} & \ms{0.2151}{0.0594} & \ms{0.1514}{0.1489} \\
Aurora & \ms{0.4423}{0.0210} & \ms{0.1551}{0.0437} & \ms{0.2162}{0.1856} \\
ClimaX & \ms{0.4151}{0.0293} & \ms{0.2129}{0.0653} & \ms{0.1587}{0.2831} \\
StormCast & \ms{0.3960}{0.0240} & \ms{0.1714}{0.0310} & \ms{0.1258}{0.1625} \\
DLWP & \ms{0.3795}{0.0274} & \ms{0.1944}{0.0807} & \ms{-0.3865}{0.2802} \\
FCN & \ms{0.4250}{0.0112} & \ms{0.1856}{0.0846} & \ms{-0.1357}{0.2571} \\
FengWu & \ms{0.4228}{0.0310} & \ms{0.1870}{0.0858} & \ms{-0.1926}{0.2194} \\
FuXi & \ms{0.4544}{0.0356} & \ms{0.2171}{0.0806} & \ms{-0.1367}{0.2885} \\
Pangu-Weather & \ms{0.3988}{0.0506} & \ms{0.1901}{0.0838} & \ms{-0.1970}{0.2216} \\
AlphaEarth & \ms{0.5276}{0.0531} & \ms{0.1782}{0.0454} & \ms{0.4639}{0.2802} \\
\bottomrule
\end{tabular}
\end{table*}

\begin{table*}[t]
\centering
\scriptsize
\setlength{\tabcolsep}{3pt}
\caption{For fixed smoke \(\mathcal{T}\) and station \(\Omega\), this table reports RMSE, MAE, and 90th-percentile absolute error on test rows with observed PM$_{2.5}\ge35$; std uses a row bootstrap over those rows. Cells report mean with small std.}
\label{tab:app_smoke_high_event}
\begin{tabular}{lccc}
\toprule
Backbone & high-smoke RMSE & high-smoke MAE & high-smoke 90th AE \\
\midrule
\ourfm & \ms{47.4870}{0.6346} & \ms{34.3954}{0.7654} & \ms{65.6213}{3.8778} \\
Prithvi-WxC & \ms{57.2224}{1.7268} & \ms{47.3871}{0.3153} & \ms{74.9666}{3.2381} \\
Aurora & \ms{57.2752}{1.7248} & \ms{47.4368}{0.3149} & \ms{75.0755}{3.1074} \\
ClimaX & \ms{57.2828}{1.7239} & \ms{47.4407}{0.3140} & \ms{75.1012}{3.0777} \\
StormCast & \ms{56.6512}{1.7517} & \ms{46.7914}{0.3281} & \ms{74.0794}{3.4707} \\
DLWP & \ms{57.0075}{1.7359} & \ms{47.1971}{0.3198} & \ms{74.4936}{3.3826} \\
FCN & \ms{57.0582}{1.7339} & \ms{47.2401}{0.3187} & \ms{74.6431}{3.1982} \\
FengWu & \ms{57.0158}{1.7357} & \ms{47.1957}{0.3194} & \ms{74.5652}{3.2871} \\
FuXi & \ms{56.9622}{1.7371} & \ms{47.1508}{0.3201} & \ms{74.3278}{3.4435} \\
Pangu-Weather & \ms{57.1282}{1.7307} & \ms{47.3050}{0.3170} & \ms{74.6830}{3.2375} \\
AlphaEarth & \ms{48.0665}{0.7904} & \ms{35.6088}{0.7341} & \ms{66.7613}{3.9235} \\
\bottomrule
\end{tabular}
\end{table*}

\begin{table*}[t]
\centering
\scriptsize
\setlength{\tabcolsep}{3pt}
\caption{For fixed heat \(\mathcal{T}\) and heat-region \(\Omega\), this table reports precision and recall for the exceedance label used by the main \(F_1\). Cells report mean with small std.}
\label{tab:app_heat_event_pr}
\begin{tabular}{lcc}
\toprule
Backbone & precision & recall \\
\midrule
\ourfm & \ms{0.9767}{0.0117} & \ms{0.9330}{0.0299} \\
Prithvi-WxC & \ms{0.8260}{0.0030} & \ms{0.9173}{0.0033} \\
Aurora & \ms{0.5920}{0.0347} & \ms{0.0517}{0.0020} \\
ClimaX & \ms{0.7397}{0.0099} & \ms{0.7994}{0.0051} \\
StormCast & \ms{0.8840}{0.0237} & \ms{0.9320}{0.0165} \\
DLWP & \ms{0.9429}{0.0085} & \ms{0.8899}{0.0167} \\
FCN & \ms{0.9408}{0.0097} & \ms{0.9111}{0.0127} \\
FengWu & \ms{0.3808}{0.2719} & \ms{0.0266}{0.0267} \\
FuXi & \ms{0.3262}{0.1262} & \ms{0.1810}{0.0481} \\
Pangu-Weather & \ms{0.1159}{0.0743} & \ms{0.0112}{0.0032} \\
AlphaEarth & \ms{0.9824}{0.0040} & \ms{0.9278}{0.0178} \\
\bottomrule
\end{tabular}
\end{table*}

\clearpage

% ============================================================
%  C  COMPARATOR ELIGIBILITY NOTES
% ============================================================
\section{Comparator Eligibility Notes}
\label{sec:comparator_audit}

All numeric comparator rows in Tables~\ref{tab:primary_results} and~\ref{tab:supporting_results}
are included only after the task form, metric, matching rule, scope, and head family are fixed.
The appendix does not repeat those full matrices.
The key eligibility rule is simple: reported rows satisfy the same contract as the row block in which they appear, while excluded rows are excluded because their representation or output form does not satisfy that contract.

\noindent\textbf{Reading rule.}
Exact-only, tolerated, union, ranking, retrieval, and regression scores answer different questions.
The fixed-contract reading is therefore to compare entries only within one row block and not to average across task forms.

\clearpage

% ============================================================
%  D  SEEDED AUDITS
% ============================================================
\section{Seeded Audits}
\label{sec:app_seeded_audits}

\subsection{Seed Robustness Summary}
\label{sec:app_seed_robustness}

Table~\ref{tab:app_seed_robustness} summarizes stochastic checks used to support the reported mean-with-std convention.
It is not a replacement for the main fixed-contract result tables.

\begin{table}[h]
\centering
\small
\setlength{\tabcolsep}{5pt}
\renewcommand{\arraystretch}{1.2}
\caption{Seed summaries for stochastic checks. Values report mean with small std over completed seeds.}
\label{tab:app_seed_robustness}
\begin{adjustbox}{max width=\textwidth}
\begin{tabular}{p{0.28\textwidth}cllp{0.18\textwidth}}
\toprule
\textbf{\(\mathcal{T}\) check} & \textbf{Seeds} & \textbf{Primary value} & \textbf{Other value(s)} & \textbf{Reading} \\
\midrule
Final burned area &
5 & log-RMSE \ms{1.1657}{0.0126} &
log-MAE \ms{1.0423}{0.0081}; Spear.\ \ms{0.6298}{0.0338} &
stable across seeds \\
Smoke PM\(_{2.5}\) &
5 & RMSE \ms{4.4646}{0.0060} &
MAE \ms{2.4108}{0.0016}; \(r\) \ms{0.6368}{0.0013} &
stable at table precision \\
Extreme heat &
5 & RMSE-C \ms{0.2179}{0.0043} &
MAE-C \ms{0.1787}{0.0018}; exceed.\ \(F_1\) \ms{0.9541}{0.0164} &
stable across seeds \\
Fire spread &
5 & exact \(F_1\) \ms{37.6700}{0.9800} &
spatial \(F_1\) \ms{80.9700}{2.0200}; AP \ms{30.0900}{1.2500} &
stable across seeds \\
Aurora paired-head check &
5 & fire-prone score diff.\ \ms{6.3500}{13.2800} &
PR-AUC and union choices differ in 2/5 seeds &
variable across seeds \\
\bottomrule
\end{tabular}
\end{adjustbox}
\end{table}

\clearpage

% ============================================================
%  E  LIGHTWEIGHT HEAD AND ADAPTATION DETAILS
% ============================================================
\section{Lightweight Head and Adaptation Details}
\label{sec:app_heads}

All frozen-transfer comparisons use the same five lightweight head architectures applied
on top of the frozen backbone representations.
Table~\ref{tab:app_head_architectures} summarises each head family, its architecture,
approximate parameter count, and the adaptation procedure used.

\begin{table}[h]
\centering
\small
\setlength{\tabcolsep}{5pt}
\renewcommand{\arraystretch}{1.3}
\caption{Lightweight head architectures used in the fixed-contract transfer comparisons.
All heads are trained from random initialisation on the frozen backbone features.
Parameter counts are approximate and depend on the feature dimensionality of each backbone.}
\label{tab:app_head_architectures}
\begin{tabular}{p{0.15\textwidth}p{0.30\textwidth}p{0.12\textwidth}p{0.33\textwidth}}
\toprule
\textbf{$\mathcal{A}$ head} & \textbf{Architecture} & \textbf{Approx.\ params} & \textbf{Notes} \\
\midrule
Constant prior &
  Outputs a fixed bias vector, ignoring input features. &
  Output dimension only &
  Provides a degenerate baseline; selected when backbone features carry no useful signal. \\
Linear probe &
  Single linear layer mapping backbone features to output. No nonlinearity. &
  $d\times c + c$ &
  Standard frozen-representation baseline. \\
Pixel MLP &
  Two-layer MLP applied independently per spatial unit. &
  $d\times h + h\times c$ &
  Captures per-pixel nonlinearity; ignores spatial context. \\
Shallow adapter &
  Two-layer MLP with a spatial context window; uses $3\times3$ convolution before the linear output. &
  $9dh + hc$ &
  Balances local spatial context with parameter efficiency. \\
Wide adapter &
  Shallow adapter with wider hidden dimension. &
  $9dH + Hc$ &
  Higher capacity variant; can overfit on small fire-event sets. \\
\bottomrule
\end{tabular}
\end{table}

\noindent\textbf{Training protocol.}
Each occupancy head-control run uses seeds $\{1,7,42,99,123\}$, the five heads listed above, and the fixed variants identity, erode-r1, and close-r1.
The spread U-Net reference is trained for 4 epochs.
The threshold $\tau$ is selected on the validation split by maximising union-$F_1$ (for occupancy) or spatial $F_1$ (for spread) and held fixed at test time.
Morphology parameters (spatial tolerance $k$, temporal tolerance $\Delta t$) are fixed as part of the evaluation contract and are not tuned after validation.

\noindent\textbf{Head selection procedure.}
For each (feature source, scope, seed) tuple, all five heads are trained independently.
The PR-AUC-based selector picks $h_R = \arg\max_{h \in \mathcal{H}} R(h)$ on the validation set;
the decision-based selector picks $h_D = \arg\max_{h \in \mathcal{H}} D(h)$ on the same set.
The selection regret $\delta = D(h_D) - D(h_R) \ge 0$ is computed on the held-out test set.

\clearpage

% ============================================================
%  F  LIMITATIONS
% ============================================================
\section{Limitations}
\label{sec:limitations}

The conclusions apply to the task forms, scopes, evaluation rules, and comparator eligibility decisions used in this study.
The evaluation covers selected wildfire decision tasks and supporting retrieval and regression task forms.
Comparator eligibility is fixed before metric values are interpreted.
This eligibility rule keeps each comparison within one task-form contract.
It also leaves some model and task pairs outside the evaluated comparison set by design.

The transfer comparison uses frozen backbones with lightweight heads.
The results therefore describe frozen-backbone transfer under the allowed head families in each contract.
Full fine-tuning, alternative adaptation procedures, and broader head families are outside the evaluated scope.
The task-specific reference baselines serve as empirical anchors for same-contract comparison.
\ourfm is a regional wildfire reference for the reported California fixed-contract experiments.


The supporting retrieval and regression checks bound the primary spatial decision claim.
They provide task-form evidence rather than a single score across all wildfire-related prediction tasks.
The analysis focuses on the reported metric families, matching rules, and fixed comparison choices.
Operational response rules, intervention costs, and deployment policies are part of wildfire early-warning use contexts~\cite{goldammer1999early,pickell2017early,farahmand2020fdeo}.
They are outside the scope of this evaluation study and are not inferred from the reported scores.

\clearpage

% ============================================================
%  G  REPRODUCIBILITY AND EVALUATION ARTIFACTS
% ============================================================
\section{Reproducibility and Evaluation Artifacts}
\label{sec:repro_compute_impact}
\subsection{External Assets and Terms of Use}
\label{sec:external_assets_terms}

We use external datasets and model assets only for research evaluation.
Access to each asset follows the original provider's portal, license, or terms of use; this submission does not imply that all assets are openly redistributable.
We do not redistribute raw external datasets, provider-hosted embeddings, or third-party model weights.
Table~\ref{tab:external_assets_licenses} records the source and terms-of-use status used to interpret reproducibility.

\begin{table}[h]
\centering
\small
\setlength{\tabcolsep}{4pt}
\renewcommand{\arraystretch}{1.18}
\caption{External assets used by the study and their source or terms-of-use status.}
\label{tab:external_assets_licenses}
\begin{tabular}{p{0.25\textwidth}p{0.34\textwidth}p{0.34\textwidth}}
\toprule
\textbf{Asset family} & \textbf{Use in this study} & \textbf{Source and terms-of-use note} \\
\midrule
NOAA HRRR fields~\cite{noaa_hrrr_ncei,noaa_hrrr_emc}
& Dynamic weather inputs for \ourfm and transfer tasks.
& NOAA provider terms and citation requirements apply. \\
NASA FIRMS~\cite{nasa_firms}
& Active-fire occupancy supervision.
& NASA Earthdata/FIRMS access terms and citation requirements apply. \\
LANDFIRE and WRC layers~\cite{landfire_fbfm40,landfire_canopy_cover,usfs_wrc_housing_density}
& Static fuel, canopy, and exposure context.
& Original geospatial-product provider terms and citations apply. \\
LandScan~\cite{ornl_landscan_2024}
& Static population context.
& ORNL/LandScan source-specific access terms apply; raw data are not redistributed. \\
WFIGS and MTBS~\cite{nifc_wfigs_perimeters,mtbs_usgs_2025}
& Event-level resources for burned-area and analog tasks.
& Original incident/perimeter-product provider terms and citations apply. \\
External Earth-FM baselines~\cite{schmude2024prithviwxc,bodnar2025aurora,nguyen2023climax,pathak2024stormcast,weyn2020dlwp,pathak2022fourcastnet,chen2023fengwu,chen2023fuxi,bi2023panguweather,brown2025alphaearth}
& Frozen comparator representations or task-model baselines.
& Original model-provider licenses and access terms apply; third-party weights are not redistributed. \\
\bottomrule
\end{tabular}
\end{table}

This note supports the NeurIPS checklist and identifies the files that support the reported claims.
This file statement does not imply full raw-data release.
The main claims can be checked from the manuscript contracts, metric
definitions, and per-head result files, even if full raw-data release is
delayed or limited. Sections~3 and~4 specify the contract components used by
the main claims: task definition, split logic, label space, tolerance
parameters, scope definitions, threshold or operating-point rules, and
lightweight-head set.

The supplementary source includes the check scripts, per-head and per-seed
CSV result files, and \LaTeX{} result tables for the expanded check and matching-rule support.
These files expose exact \(F_1\),
tolerated \(F_1\), union-\(F_1\), PR-AUC, per-head selection,
top-1 agreement, and selection-regret arithmetic. The manuscript also includes
full figure and table reproduction values in result tables and appendix tables.
These files provide a runnable check of the
selection-regret arithmetic and the table-construction logic from fixed
per-head rows. The seeded occupancy check uses seeds
$\{1,7,42,99,123\}$, and the spread task-specific U-Net check uses repeated seeds; reported error bars are standard deviations over the completed
seeded runs. Full raw wildfire inputs and large feature arrays are not
released at submission because redistribution and storage constraints require a
separate review.

For stochastic results, the paper reports mean with standard deviation over repeated seeds.
For fixed-output or fixed-feature controls, the table uses one fixed output or feature set; the changed item is the matching rule or selection metric.

The reported experiments use two resource classes on a shared Slurm-managed
cluster. Tabular retrieval/regression checks and same-feature head controls run
on CPU workers with 4 to 8 cores, 24 to 64~GB host memory, and 2 to 4~hour wall-clock
limits. Spread U-Net training and threshold calibration run on single-GPU jobs
with one B200 GPU, 8 CPU cores, 96~GB host memory, and a 4~hour wall-clock
limit. The seed/check waves reported in the appendix correspond to roughly
78 CPU job-hours and 12 GPU job-hours of scheduled wall-clock budget;
exploratory runs are not included in the reported compute accounting.

The raw-data limitation is separate from the selection-regret files.
The supplementary source is sufficient to inspect the selection-regret arithmetic and reproduce the reported tables.
Full end-to-end recomputation from raw wildfire inputs is not included at submission because redistribution review is still required.
The broader impact is evaluation-facing rather than operational.
Better reading of wildfire transfer evidence can reduce overconfident benchmark claims, while misread transfer results could still encourage inappropriate reliance on models with low decision scores.
For that reason, the paper keeps its claims wildfire-centered, decision-task
specific, and explicitly separate from any predictive deployment
recommendation.