Per-redshift-bin RMSE and η_0.15 on the DR6 test set at n_train = 121,626 for the three methods with the lowest overall unweighted RMSE in Table 3 (top panel): TabPFN 2.5, MDN-deep, and GBM-Point. Bins have the fixed width Δz_spec = 0.5; bins with fewer than 30 test objects are dropped. Shaded bands are ±1 SE. All three methods agree to within one SE in the midredshift range z_spec ∈ [0.5, 3.0]; differences are concentrated at the low- and high-z ends, where the foundation method performs the best.