In multi-turn reinforcement learning, your custom reward function decides what the model actually learns. This post shows how to design a composite multi-turn reward for Amazon Nova Forge, execute model-generated code safely inside it, and instrument each component to catch the pitfalls that quietly collapse a reward.