In aerospace engineering, many systems exhibit strongly nonlinear, divergent, or unstable behavior when driven outside their nominal operating regimes: Their control design has to rely upon conservative assumptions and dedicated envelope protection logics restricting operation within regions with substantially linear dynamics. This often leads to a laborious, expensive, and timeconsuming development process. In this paper, we propose a reinforcement-learning-inspired technique for envelope protection aimed explicitly at reducing the required designer workload by removing the requirement to explicitly define envelope limits. Our method comprises three key elements: (1) Neural nets for data-only dynamic identification, (2) a novel method for exploiting the neural model's parametric uncertainty to generate gradients that contain the dynamics within the known, safe envelope and, finally, (3) a Temporal Backpropagation (TB) calculation which converts the resulting optimal control problem to one of training of a deep, recurrent neural net architecture. The method's performance is assessed in simulated experiments: A simple numerical case demonstrates the algorithm's key characteristics. Finally, testing on a 6-DoF aircraft simulator evaluates the effectiveness of a TB-based Flight Envelope Protection (FEP) in flight sequences including aggressive pilot inputs and highly nonlinear post-stall aerodynamics.