My understanding is that for each calculation it uses first the most recent, then if there are multiple most recent it uses the highest resolution data available for a given time step.
This means that if you have a something like a very new global model and an older meso scale model you can wind up with "interesting" results as data points for a given area will be a different models due to diff mixing models for given time steps as meso scale models tend to have shorter timesteps than global models.