Relevance sampling runs a configurable sample count of repeated yes/no evaluation calls per item across all active models to capture stochastic variance in model outputs.