Specification that maps states or actions to scalar rewards in reinforcement learning, defining the agent’s objective, often the hardest part to get right.