Data as of Sep 6, 2026, 12:26 AM UTC

Bias Prevention & Fairness

HRMLESS evaluates every AI system for potential bias before deployment and continuously monitors fairness metrics in production. Our approach is grounded in the NIST AI RMF Measure function and aligned with EEOC guidelines for employment-related AI systems.

Production Fairness Attestation

PASS

320 records scored by the live NervAI Grading Engine using controlled counterfactual evaluation. Identical interview responses were scored across every combination of gender, race/ethnicity, and age -- any score difference would be direct evidence of bias.

Attribute	Dem. Parity	KS p-value	Competency Parity
Gender	1.0 PASS	1.0 PASS	0.019 PASS
Race / Ethnicity	1.0 PASS	1.0 PASS	0.048 PASS
Age	1.0 PASS	0.998 PASS	0.019 PASS

Attestation date: July 9, 2026 · System: NervAI Grading Engine · 320 records

Download Full Report

NYC Local Law 144 Compliance

COMPLIANT

The NervAI Grading Engine has been independently audited for bias pursuant to NYC Administrative Code §§ 20-870 through 20-874 (Local Law 144 of 2021). Selection rates and impact ratios were calculated across sex, race/ethnicity, and intersectional categories per EEOC Uniform Guidelines (29 C.F.R. § 1607.4).

Analysis	Categories	Impact Ratios	Adverse Impact
Sex Categories	Male, Female	1.0 across all	None Detected
Race/Ethnicity	White, Black, Hispanic, Asian	1.0 across all	None Detected
Intersectional	8 categories (sex × race/ethnicity)	1.0 across all	None Detected

Audit date: July 9, 2026 · Standard: Four-fifths rule (impact ratio ≥ 0.80) · 320 assessments · Selection threshold: score ≥ 0.50

Testing Methodology

Our five-phase approach to bias prevention ensures systematic coverage from initial assessment through ongoing production monitoring.

Identify

Catalog protected characteristics relevant to each AI system's context of use. Define the population segments and decision boundaries that require fairness evaluation.

Measure

Apply quantitative fairness metrics to model outputs across demographic groups. Compare observed disparities against defined thresholds to flag potential bias.

Evaluate

Assess flagged disparities in context. Determine whether observed differences indicate actionable bias or reflect legitimate, job-related distinctions.

Remediate

Implement corrective actions when bias is confirmed. Re-test after remediation to verify the issue is resolved without introducing new disparities.

Monitor

Continuously track fairness metrics in production. Detect drift in bias indicators and trigger re-evaluation when thresholds are exceeded.

Fairness Metrics

We track multiple complementary metrics to capture different dimensions of fairness. No single metric is sufficient; together they provide a comprehensive view of AI system equity.

Demographic Parity

EEOC Four-Fifths Rule

Selection rates across demographic groups should be comparable. A ratio below 0.8 (four-fifths rule) triggers review.

Equalized Odds

NIST AI RMF MS-2.6

True positive and false positive rates should be consistent across groups, ensuring the system is equally accurate for all populations.

Calibration

NIST AI RMF MS-2.7

When the system assigns a score or probability, that score should mean the same thing regardless of group membership.

Predictive Parity

Industry Best Practice

The precision of the model (positive predictive value) should be comparable across groups.

Continuous Monitoring

Automated bias metric computation on every model update and at regular production intervals
Statistical drift detection comparing current metric values against deployment baselines
Threshold-based alerting when any fairness metric exceeds predefined limits
Quarterly comprehensive bias audits with full demographic breakdowns
Annual third-party bias assessment readiness for jurisdictions that require independent review

Remediation Process

When a fairness concern is identified, we follow a structured remediation process to ensure timely and effective resolution.

Detection

Automated monitoring or manual review identifies a potential fairness concern.

Triage

AI governance team assesses severity, affected populations, and business impact within 48 hours.

Root Cause

Investigation determines whether the disparity stems from training data, model architecture, or feature selection.

Correction

Implement targeted fixes -- data rebalancing, feature engineering, threshold adjustment, or model retraining.

Validation

Re-run full bias test suite to confirm remediation effectiveness without regression.

Documentation

Record findings, actions, and outcomes in the AI Risk Register and update relevant Model Card.

AI System Fairness Status

Current fairness evaluation status for published AI systems.

- NervAI Grading Engine

Model Card published Jun 30, 2026

Evaluated

- NervAI Interview Engine

Model Card published Jun 30, 2026

Evaluated