Best I can tell, they are using the ML model to detect the errors. Isn't this a bit of an ouroboros? The model will naturally get better, because you are only c

Best I can tell, they are using the ML model to detect the errors. Isn't this a bit of an ouroboros? The model will naturally get better, because you are only correcting problems where it was right but the label was wrong. It's not necessarily a representation of a better model, but just of a better testing set.