A reward system that breaks complex requests into atomic, checkable questions to provide interpretable feedback for model training.