Great write-up! I think model monitoring and retraining would be an interesting follow-up topic. How would you think about a system like this holistically and determine when a model needs retraining v. when a routing rule simply needs to be updated to choose an existing model?