- The marginal four-fifths test a vendor self-audit ships passes here: race 0.8196, sex 0.8744, both over the 0.80 line. The intersectional race x sex test NYC Local Law 144 actually requires fails at 0.6471 for Black/Female. Same synthetic 1,040-candidate export 🧵
- Marginal means one attribute at a time. Race on its own, sex on its own. Intersectional means the cells. LL144 asks for the cells. A hiring stack can clear every marginal line and still be cutting one specific cohort out of the funnel.
- The arithmetic on that synthetic export: White/Male advances 68 of 130, 52.31%, the reference cell at 1.00. Black/Female advances 44 of 130, 33.85%. Divide and you land at 0.6471, well under the 0.80 threshold the marginal test just cleared.
- A second cell misses too. Hispanic/Female sits at 0.7647. Under Benjamini-Hochberg FDR control at alpha 0.05, only Black/Female is statistically robust (q = 0.0185). Hispanic/Female (q = 0.1629) is flagged and explicitly not called significant.
- That separation matters as much as the failure does. An audit that books every flagged cell as a finding is about as useful as one that books none. The engine sorts a robust result from a chance one before either reaches a jurisdiction deliverable.
- So we put all of it in deterministic code outside the agent framework. numpy computes the ratios, rule packs compare them to each regime's threshold, and a policy gate sets PASS, FAIL, NEEDS PROOF or CONFLICT. The agent crew narrates and cannot move a number.
- Which is why the figures do not drift between runs. Same seeded input, same output. 12 of 12 unit tests pin this exact ground truth: the marginal ratios pass while the intersectional ones fail, run after run.
- The Black/Female gap is planted, deliberately, in synthetic data. It says nothing about any real employer. It is there to answer one question: would a system that reports only marginal ratios ever surface it? It would not.
- Across 13 obligations in six regimes, this export returned 2 PASS, 2 FAIL, 7 NEEDS PROOF and 2 CONFLICT. Nine of those 13 landed on a named human rather than auto-cleared. One universal fairness score has nowhere to put a result shaped like that.
- For compliance and HR-tech folks: open the last bias audit PDF a vendor sent you. Does it print the race x sex cells, or only the two marginal ratios? Which one you received decides whether you have seen your own funnel. #LL144 #AIGovernance #HRTech #EmploymentLaw
- Clarion is a demo on seeded synthetic data, not a deployment. Vendor connectors are stubbed fixture adapters. We build the pre-audit package; an independent auditor signs it, not us. See how it works: https://veriprajna.com/demos/ai-hiring-compliance
Published on X · September 11, 2026
On social media
See this post on its original platform
In our archive