This is great stuff. I’m not a data scientist, or even a numbers nerd, so I’m probably the worst person to ask for input, but instead of giving concrete suggestions, I’m going to pile onto what you offered and suggest that, instead of going with this or that methodology, perhaps you could come up with a way to measure for ALL indicators, weighing and normalizing against a scale of optimal equality indicators.
It still wouldn’t be perfect, but it would solve the Pocahontas problem you mention, as well as the Bechdel testing. Essentially, you could account for outliers and edge cases while still meeting some kind of approachable universal means of equality assessment.
If I think of more concrete means of doing so, I’ll certainly weigh in!