I have a better model to decrease disparities - just always return False. Equality solved.

But I analyzed the data from their github and it seems like it actually improved total model accuracy in most cases. But I had to dig really deep to find it out, unless I missed something in the paper.