Training labels are used only by a Euclidean coordinate-only nearest-neighbour classifier. Ties use the smallest row-major training index. Patches are audited geometrically but never supplied to this learner. Score differences cannot be attributed causally to patch overlap. Smoothing changes the generating field; changing patch width changes which interior centres are eligible. At fixed settings the random and blocked rows use identical field values, class labels and training counts before buffering, but different training locations and test populations.

Buffer b excludes a test centre if its Chebyshev distance to any training centre is ≤b. Setting b=p−1 eliminates all shared raw support for equal odd-width p patches. It does not eliminate longer-range dependence. No padding, fitted feature transforms, spectral data, hyperparameter selection or test-label-based training is used. Moran’s I uses unit weights for both directions of horizontal/vertical adjacency on the scalar field. It is descriptive; no p-value is calculated. Next seed varies the generated scene and random split together, not only optimiser initialisation.