Thanks, I didn’t know the term. I think that’s exactly what we are doing. Our transformation creates a subgraph of the rvs that are conditionally dependent on the marginalized variable with inputs being the set of rvs that are inputs to the marginalized variable + dependent variables.
The logp is then computed for this subgraph alone via enumeration/qmc.
It’s very trivial for us to find this partition (Markov blankets?) from the graph representation of PyMC models