We consider layered control architectures where a constraint-enforcing upper layer is cascaded with a lower layer controlled actuator. As we aim for a design where each layer requires as little knowledge as possible of the other, the upper layer is based on a model that neglects the lower layer dynamics, and includes instead additive uncertainty. The uncertainty set is constructed and “declared” by the lower layer based only on constraints on the command “declared” by the upper layer. This results in a contract between upper layer and lower layer guaranteeing a bound on the prediction error if the command satisfies the declared constraints. The command and plant constraints are robustly enforced by model predictive control with a robust control invariant set. The stability properties are analyzed, and a case study of vehicle steering control is shown.