K-BENCHMARK Methodology
An open, transparent framework for evaluating AI model trustworthiness.
What is K-BENCHMARK?
K-BENCHMARK is ALPAR AI's open methodology for scoring AI models across safety, truthfulness, fairness, privacy, robustness, and transparency dimensions. Scores are computed from verified incident reports, cross-audit engine results, and domain expert evaluations using Wilson-score confidence intervals.
Evaluation Categories
Safety
Harmful content generation, jailbreak resistance, and refusal behavior under adversarial prompts.
Truthfulness
Factual accuracy, hallucination rate, and citation fidelity across knowledge domains.
Fairness
Bias in outputs across demographic groups, representation, and equitable treatment.
Privacy
Training data memorization, PII leakage, and membership inference resistance.
Robustness
Performance under distribution shift, adversarial inputs, and edge cases.
Transparency
Model documentation quality, disclosure practices, and auditability.
Scoring Method
Each category is scored using a Wilson-score confidence interval, which accounts for sample size and provides a statistically rigorous lower bound. This prevents small-sample noise from inflating ratings.
Wilson score = (p + z²/2n - z√(p(1-p)/n + z²/4n²)) / (1 + z²/n)
Wilson-score is a standard method in binomial proportion estimation, commonly used in platforms like Reddit and IMDb for robust rating systems.
Pipeline
Incident Submission
PII Guard
Cross-Audit Engine
Wilson Scoring
Expert Review
Data Sources
Scores are derived from: (1) verified incident reports submitted via the ALPAR platform, (2) automated cross-audit engine evaluations using established benchmarks (MMLU, GSM8K, IFEval, BBH), (3) domain expert assessments from the L3 advisor network, and (4) publicly available model documentation and technical reports.
K-BENCHMARK scores are informational and provided 'as-is'. They do not constitute a compliance certification or legal opinion. ALPAR AI is 'AI Act ready/aligned', not compliant. See Terms of Service for full disclaimer.