Fascinating research, which incidentally, is verifiable simply by copying the sentence instances and trying it yourself. What is certainly the case, is that labelling and output restrictions are not entirely dependent on the input but on a hidden taxonomy of identity groups. We should all see the obvious problems with this
