Some kind of circular review process is needed

AKA a human in the loop.

Currently, the process ingests labelled training data files, but we need to be able to correct the bad predictions, then retrain a better model on that.

I wonder if the mixfile would be a good foundation for this?