Quantization reconfigures decision-boundary geometry independent of accuracy

- Under quantization, a classifier's rasterized decision-boundary mask and multiclass junction topology measurably reconfigure even when top-1 accuracy is preserved or improved. [kiseleva-2026-boundary-aware-quantization] - On a one-hidden-layer digits network, boundary-mask Jaccard rises 0.428 (8-bit) to 0.970 (4-bit) to 0.986 (2-bit) while 4-bit accuracy stays at 0.9733, geometry moving independently of accuracy. [kiseleva-2026-boundary-aware-quantization] - On the CIFAR-10 residual CNN, 6-bit weight quantization changes 5.3% of held-out decisions and 24.5% of low-margin boundary-band decisions for only -0.0029 accuracy; calibration-set boundary Jaccard predicts held-out Jaccard (r=0.947-0.994). [kiseleva-2026-boundary-aware-quantization] - Correlational/diagnostic, measured on 2D grid/PCA slices; the residual-CNN architecture is left unspecified in the paper. [kiseleva-2026-boundary-aware-quantization]