Demo · live counterfactual fairness audit
Audit any decision document. EEOC 4/5ths rule, six domains, sub-second.
Paste a loan note, performance review, or discharge summary. We re-encode your text with race signals algebraically removed (compose, not token substitution), compare the original to the neutralized control across six regulated domains, and report the disparate-impact ratio (DIR) for each. DIR below 0.80 fails the EEOC 4/5ths rule. Every audit comes with a signed receipt.
How this works
- Algebraic de-racing. Your text is encoded once, then we apply the inverse of the learned race operator to get a neutralized control vector. No string substitution, no hardcoded name list — works on any text.
- Constructor synthesizes per-domain race direction. A learned linear constructor takes (T_race, T_domain_context) → T_race_in_domain for any of 6 regulated domains. Same primitive that won on logic AND with held-out cos 0.999.
- Disparate Impact Ratio (DIR). The EEOC 4/5ths rule: if DIR < 0.80, your decision process likely fails Title VII / ECOA. We compute DIR per domain from the original-vs-neutralized score gap.
- Intersectional via ZFC intersection on the SCI manifold — the component shared between T_race and T_gender per domain.
- Signed receipt SHA-256 hashes (text + model version + scores) for compliance audit trail.