Bias testing report

Scope of testing

We periodically re-score held-out synthetic and consented evaluation sets while holding job requirements constant. Tests look at score gaps across gender-coded names, ethnicity-proxy name lists, and university prestige signals that are not job-relevant.

Latest results (evaluation set)

On the current evaluation set, mean overall-fit difference between gender-coded name pairs was under 1.5 percentage points when experience and skills were identical. University-name ablation (replacing elite vs regional schools with a generic label) changed shortlist membership for fewer than 3% of otherwise identical profiles.

These figures describe the matching ranker on evaluation data. They are not a legal certification and do not replace customer-specific impact assessments.

What we do with findings

If a test exceeds internal thresholds, ranking features that encode protected or proxy attributes are disabled or reweighted before the next production rollout. Recruiter-facing explanations must cite job-relevant evidence (skills, tenure, domain), not demographic or school-prestige language.

Request the full packet

Customers under NDA can request the latest methodology appendix, slice tables, and change log from [email protected]. The public summary on this page is updated when a new evaluation cycle completes.