AI Fairness & Bias Testing Methodology

• Assess on the job, not the person. Candidates are scored only on criteria the employer defines as relevant to the role.
• A human always decides. Nuvett ranks and informs; it never makes an autonomous hiring decision.
• Test, don’t assume. We measure outcomes for disparity rather than presuming the model is fair.
• Publish the method. Employers and candidates can read exactly how we test — this document.

Our primary fairness metric is the Adverse Impact Ratio (AIR), the basis of the US “four-fifths rule” and a widely accepted test of selection fairness.

The “selection rate” is the share of a group that passes the assessment threshold the employer is using for that role. Where candidate volume is large enough, we supplement the four-fifths test with a statistical-significance test (standard-deviation analysis / p < 0.05), because ratios alone can mislead on small numbers.

Which characteristics we test depends on the employer’s jurisdiction. We test the grounds protected by the law that applies to that hire.

Testing for a characteristic requires data about it. We collect demographic data only where a candidate provides it voluntarily, hold it separately from assessment scoring, and use it solely to run these fairness checks — consistent with the NDPA and GDPR treatment of sensitive data. Where a group is too small to test reliably, we say so rather than draw a false conclusion.

• Minimum sample: we do not compute an AIR for any group with fewer than ~30 candidates; results below that are treated as indicative only.
• Proxy check: we review that scoring criteria don’t act as stand-ins for a protected trait — for example school attended, home address, or name signalling ethnicity or origin.
• Cadence: fairness is evaluated per role when the role closes, and aggregated across roles on a quarterly basis to catch patterns a single role would hide.
• Records: test inputs, results, and any action taken are logged, so the analysis can be reviewed later by the employer or a regulator.

When a test flags a problem — an AIR below 0.80, or a statistically significant disparity — we act rather than note it:

1. 1Investigate which criterion or question is driving the disparity.
2. 2Re-examine and re-weight the criteria, or remove a component found to be a biased proxy.
3. 3Flag or pause the role’s configuration where a fair result can’t be assured, and alert the employer.
4. 4Document the finding and the remedy, and re-test after the change.

• The employer makes the final hiring decision on every candidate. Nuvett provides ranked, explained results — it does not select or reject autonomously.
• Assessment outputs are explainable: a score can be traced to the criteria and responses behind it, so a human can review why a candidate ranked where they did.
• Candidates are told an AI system is used in their assessment, and may request review of a result.

We publish this to be trusted, so we’re honest about the limits. Fairness testing depends on demographic data that is voluntary and sometimes sparse, and no statistical test can guarantee that every individual outcome is fair. What we commit to is a defined, repeatable process that surfaces disparity, acts on it, and improves over time. This methodology is reviewed at least every six months and updated as law, standards, and our own learning evolve.