Naive fairness tests flip sex: Male → Female and call it done.
They miss the proxies. relationship=Husband still screams Male — your model reads the row as off-distribution and absorbs the bias silently.
"relfair" propagates the intervention through the causal graph (Husband→Wife, occupation, household role) so counterfactuals stay on the manifold.
Result: 3–4× more discrimination detected on Adult (+17pp), ACS Income (+27pp), German Credit (+8pp).
Since I’ve been asked a lot:
Buy stock in several companies that make products & services that *you* believe in.
Only sell if you think their products & services are trending worse. Don’t panic when the market does.
This will serve you well in the long-term.