C4 Symmetry Pocket tests architectural rotation invariance against two matched controls: canonical-only training and explicit rotation augmentation. Every variant uses the same base CNN and parameter count. The invariant network averages predictions across the four quarter-turn group elements.

The benchmark reports accuracy at 0, 90, 180, and 270 degrees, plus the worst angle. Group averaging requires four base-network evaluations at inference, so the robustness result is not presented as free computation.

All three variants contain exactly 1,201 parameters. Canonical-only training fell to 38.90% worst-angle accuracy, rotation augmentation reached 99.90%, and C4 group averaging reached 100% at every tested quarter turn. The invariant model requires four base-CNN evaluations per prediction.