Using this sampling-based approach you get correct covariance modeling for free. You have to only sample leaf values that are used in multiple places once per evaluation, but it looks like they do just that: https://github.com/mattt/Uncertain/blob/962d4cc802a2b179685d...