DiCo

Replacing full SND with sparse Graph-SND preserved closed-loop diversity control in VMAS Dispersion experiments. The DiCo head-to-head comparison reported metric speedups of about 9.24x to 9.63x with a small paired reward difference. At n=50, Bernoulli-0.1 Graph-SND used 122.5 expected sampled edges compared with 1,225 full pairs. At n=10, several sparse Graph-SND variants tracked the desired SND level and matched the full-SND controller's reward within seed variation. In DiCo experiments, empty Bernoulli draws fall back to the full pair set so the control ratio remains defined.