Services — Minerva AI Assurance

Core Service

Bias Testing for AI Decision Systems

We analyze AI outcomes across demographic groups and intersections to identify adverse impact — in hiring, lending, insurance, risk scoring, and any context where AI decisions affect people. Selection rates, impact ratios, and error-rate analysis with reproducible methodology and practical remediation guidance.

Includes AEDT compliance audits for automated hiring and promotion tools, structured for NYC Local Law 144 public disclosure requirements.

Statistical adverse impact analysis and documentation. Not legal advice; not a guarantee of regulatory outcomes.

Governance

AI Governance Readiness Review

A plain-English assessment of your AI governance posture — policies, roles, documentation, accountability structures. You receive a findings report, risk-ranked gap analysis, and a 90-day remediation roadmap aligned to the framework that matters most for your context.

What You Get

Evidence that holds up within the agreed framework.

Our audit reports are designed for the people who need to act on them — compliance officers, risk committees, and boards. Clear findings, reproducible methodology.

What an Audit Includes

✓Demographic group identification and sample size validation

✓Selection rate calculation per group

✓Impact ratio analysis with four-fifths rule assessment

✓Intersectional analysis (e.g., sex × race/ethnicity)

✓Small-sample exclusion documentation with rationale

✓Risk-ranked findings with remediation guidance

✓Complete methodology disclosure and reproducibility artifacts

✓Limitations and scope disclaimers

Sample Output — Gender AnalysisIllustrative

Group	Applicants	Above-Threshold Rate	Impact Ratio	Result
Male	1,291	51.4%	1.00	Pass
Female	1,209	48.5%	0.94	Pass

Illustrative only. Example threshold: IR ≥ 0.80 (context-dependent). Thresholds and interpretation depend on sample sizes and applicable regulatory framework.

Anonymized Example

AI Hiring Screen — Adverse Impact Detection

System: AI-generated candidate scoring model used to screen 2,500+ applicants across five job categories.

Finding: Single-attribute analysis showed all race/ethnicity groups above the 0.80 threshold except one (IR = 0.76). Intersectional analysis revealed compounding: the intersection of that group with female gender produced an IR of 0.69 — significantly below threshold and invisible in single-attribute testing alone.

Outcome: Board-ready report delivered in 12 business days with risk-ranked findings and specific recommendations for model recalibration and monitoring cadence.

Methodology

How we test.

Metrics We Use

Selection rate / above-threshold rate per group
Adverse impact ratio (four-fifths rule)
True positive / false positive rate parity
Intersectional group analysis
Sample size validation and small-group handling

Metrics are chosen with the client based on context, applicable policy, and the specific decision system under review.

Frameworks We Align To

Singapore FEAT Principles (MAS)
AI Verify Testing Framework (IMDA)
NIST AI Risk Management Framework
ISO/IEC 42001 AI Management Systems
NYC Local Law 144 (AEDT)
EEOC Uniform Guidelines (Four-Fifths Rule)

Independent assessment. Practical evidence.