 Solution

The Rao–Blackwell theorem says that for convex loss, conditioning any estimator on a sufficient statistic cannot increase risk; for squared error it preserves the mean and weakly reduces variance.