Can’t see how this doesn’t defeat the purpose, if you have a mock data generator of sufficient fidelity you already have a well-defined mechanism to describe the data, shouldn’t you be using that instead of developing a new model to capture the characteristics of the model that creates those characteristics?