Location location location - does a tabular foundation model need spatial features

Mass appraisal models encode the dependence of property prices on location explicitly, through spatial weights matrices, spatial lags, geographically weighted local models or engineered neighbourhood features. We ask whether a tabular foundation model can do without them. TabPFN-3.5 receives the property attributes and the geographic coordinates as two ordinary numeric columns, with no spatial modelling of any kind, and is compared with an OLS hedonic model, a spatial autoregressive lag model (SAR, GM estimation), XGBoost with a k-nearest-neighbour spatial lag, rotated coordinates and nested tuning, and a geographically weighted XGBoost. The data are 82,187 declared transaction prices from the São Paulo real-estate transfer tax (ITBI) in 2025 and 47,810 in the rst seven months of 2026, for apartments, houses and commercial units. The evaluation has two legs: leave-oneblock-out cross-validation over ten spatial blocks, and an out-of-time test that fits on 2025 and predicts 2026. Refutation criteria were fixed in advance: worse pooled error, losses in most blocks under paired tests, or more residual spatial autocorrelation (Moran's I) than the explicitly spatial baselines. Out of time, the plain-table model beats all three spatial specialists in at least 9 of 10 blocks (RMSE 0.265 in log unit price against 0.283-0.326; MAPE 19.5% against 20.925.0%) and leaves the least residual autocorrelation of the models compared (I = 0.092); with the raw postal code in place of the coordinates one text column, no geometry every one of those verdicts holds. Across unseen blocks it ties both XGBoost variants on error and beats SAR, but leaves more spatial structure in its residuals (I = 0.348 against 0.309-0.326), a partial refutation, on the record. Permutation SHAP shows the two coordinates as the model's strongest inputs; the postal-code result shows that what they carry, inside an observed market, is location as an identier rather than geometry, and that only coordinates extrapolate to districts without sales. For mass appraisal inside a market with a sales history, a foundation model given the raw table is the strongest predictor in this comparison, with one fit of fifteen seconds and no spatial engineering.