MARIA: A Framework for Marginal Risk Assessment without Ground Truth in AI Systems