Fair by architecture. Audited by outsiders. Monitored on every requisition.
Most vendors publish a fairness promise. We publish the mechanism, the monitoring, and the auditors. This page is written for your legal team — send it to them.
The MERIT-FAIR™ assurance framework
Content-only scoring
Scoring models receive transcript content only — never video frames, audio features, names, postcodes, or inferred demographics. Emotion recognition is not a disabled feature; it is an absent capability, which is what the EU AI Act Art. 5 actually requires of hiring systems.
Policy compiler
Hiring rules are compiled, not free-typed. Criteria that reference or proxy protected attributes are refused at configuration time with a lawful reformulation suggested — discrimination is blocked before an interview ever runs.
Per-requisition impact monitoring
Selection rates are monitored against the 4/5ths rule per requisition with alerts before shortlists ship, not in a quarterly report after the harm. Every alert and its resolution is retained in the audit pack.
Evidence or abstention
Every competency score links to verbatim, timestamped interview evidence. When evidence is insufficient, agents abstain and the system routes a follow-up probe — uncertainty is surfaced, never papered over.
Independent disparate-impact audit
A third-party statistical audit of MERIT-8 scoring outcomes across gender, age band and language background, following the methodology standard set by NYC Local Law 144 bias audits. Summary results will be published here; full report available to customers under NDA.
ISO/IEC 42001 — AI management
Certification of our AI management system: model change control, incident response, human-oversight procedures, and supplier (model-provider) governance. SOC 2 Type I runs in parallel.
Model cards & rubric versioning
Every requisition records the rubric version, agent prompt versions and model versions used. Change a rubric and the audit trail shows who, when, and which candidates were scored under which version.
Human decision, named reviewer
No candidate is rejected by AI alone. Every disposition carries a named human reviewer and an override trail — the sentence that ends most legal reviews.
Jurisdiction readiness
| Regime | What it demands | Our posture |
|---|---|---|
| EU AI Act (high-risk: employment) | No emotion inference; human oversight; logging; transparency | Emotion inference architecturally absent · named-reviewer flow · full event logs · candidate AI-use notice |
| NYC Local Law 144 | Annual independent bias audit; candidate notice | Audit commissioned (Q3 2026) · notice built into invite & consent flows |
| Illinois AIVIA / HB 3773 | Consent, explanation, deletion; no zip-code proxies | Consent-gated interviews · deletion on request · postcode proxies blocked by policy compiler |
| Australia (AHRC guidance, Privacy Act) | Fairness, transparency, APP compliance | AU data residency option · content-only scoring · candidate feedback for every applicant |
REQUEST THE AUDIT METHODOLOGY PACK OR OUR DPIA TEMPLATE: [email protected] · CANDIDATES CAN REQUEST THEIR INTERVIEW DATA AT ANY TIME