\begin{table}[htbp]
\centering\scriptsize
\caption{Task B ablations at the municipality level: MAE (×100) and difference to the reference variant (per-origin mean, * = 95\% stationary-bootstrap interval excludes zero)}\label{tab:ablation_b}
\resizebox{\textwidth}{!}{%
\begin{tabular}{lrrrrrr}
\toprule
Variant & h=1: MAE & vs ref & h=3: MAE & vs ref & h=6: MAE & vs ref \\
\midrule
DSGFN (full) & 2.66 & ref & 4.12 & ref & 5.05 & ref \\
DSGFN without distance kernel & 2.64 & -0.02 & 3.95 & -0.17 & 5.08 & +0.02 \\
DSGFN without co-movement & 2.70 & +0.04 & 4.01 & -0.11 & 4.87 & -0.19 \\
DSGFN with uninformative prior graph & 2.60 & -0.06 & 3.92 & -0.20 & 4.46 & -0.59* \\
QANSS (full) & 2.47 & ref & 3.75 & ref & 4.80 & ref \\
QANSS without parent drifts & 2.52 & +0.05 & 3.73 & -0.02 & 4.82 & +0.02 \\
QREFA (NNLS stack) & 2.59 & ref & 3.75 & ref & 4.68 & ref \\
Equal-weight combination of the four & 2.38 & -0.21* & 3.45 & -0.30 & 4.30 & -0.38 \\
QREFA without QANSS & 2.54 & -0.05 & 3.73 & -0.02 & 4.95 & +0.26* \\
QREFA without BHPF & 2.53 & -0.05 & 3.88 & +0.13 & 5.04 & +0.36 \\
QREFA without DSGFN & 2.43 & -0.16 & 3.64 & -0.11 & 4.94 & +0.26 \\
QREFA without RAREF & 2.62 & +0.03 & 3.65 & -0.10 & 4.64 & -0.04 \\
Ridge panel (full features) & 2.54 & ref & 3.73 & ref & 5.00 & ref \\
Ridge without own momentum & 2.65 & +0.10* & 4.03 & +0.30* & 5.54 & +0.53 \\
Ridge without parent (region/province) features & 2.57 & +0.02 & 3.78 & +0.05 & 4.89 & -0.11 \\
Ridge without volume features & 2.57 & +0.03 & 3.76 & +0.03 & 5.11 & +0.10 \\
\bottomrule
\end{tabular}}
\par\vspace{3pt}\parbox{\textwidth}{\footnotesize Reference variants: DSGFN, QANSS, QREFA and the full ridge panel. Positive differences mean the ablated variant is worse.}
\end{table}