The idea is to train a very small model, perhaps 7 or 8 billion parameters, on your actual usage so that it can run at the edge. It learns how you use models and assigns the best verifier, planner and executor for each turn.