@jornfranke Maybe that technique with the BIG assumption they’re open about the training data - would be the best way to test this?