In this paper we consider large scale distributed committee machines where no local data exchange is possible between neural network modules. Regularization neural networks are used for both the modules as well as the combiner committee in an embedded architecture. After the committee training no module will know anything else except its own local data. This privacy preserving obligation is a challenging problem for trainable combiners but crucial in real world applications. Only classifiers in the form of binaries or agents can be sent to others to validate their local data and sent back average classification rates. From this fully distributed and privacy preserving mutual validation a coarse-grained matrix can be formed to map all members. We demonstrate that it is possible to fully exploit this mutual validation matrix to efficiently train another regularization network as a meta learner combiner for the committee.