Evaluation Framework
Seven Critical Safety Pillars
Every RateAI evaluation measures performance across seven independently scored pillars. No single strong score can compensate for a critical failure — our downgrade engine enforces accountability at every layer.
RateAI's evaluation methodology is inspired by international standards in AI governance, including ISO/IEC 42001, the EU AI Act, and NIST AI RMF. Pillar weights are calibrated to reflect real-world risk exposure and can be adjusted for domain-specific certification profiles.
20%
🏛️
Pillar G
Governance
Evaluates policy adherence, documentation completeness, access controls, audit logging, and organizational accountability structures for the AI system.
15%
🎯
Pillar H
Hallucination Control
Measures factual accuracy, source grounding, confabulation rates, and the model's ability to express calibrated uncertainty rather than fabricating confident answers.
15%
🔒
Pillar D
Data Privacy
Tests PII leakage, training data memorization, prompt injection susceptibility, and compliance with GDPR, CCPA, and other data protection frameworks.
15%
⚖️
Pillar E
Ethics & Bias
Assesses demographic bias, representational fairness, harmful content generation, value alignment, and the model's handling of sensitive social topics.
15%
🛡️
Pillar R
Robustness
Stress-tests adversarial input resilience, jailbreak resistance, performance under distribution shift, and reliability degradation across edge cases.
10%
📋
Pillar C
Regulatory Compliance
Verifies alignment with applicable regulations including the EU AI Act, NIST AI RMF, ISO/IEC 42001, HIPAA (for healthcare AI), and emerging national AI standards.
10%
🔍
Pillar X
Explainability
Evaluates whether the model can meaningfully explain its outputs, provide reasoning traces, and support human oversight — critical for high-stakes deployment contexts.
