The freshness-based diversity weight exponentially deprioritises heavily tested classes by weighting selection via probabilistic sampling proportional to each class's inverse round count raised to a configurable power.