This study presents a safe, frequency-separated hybrid control-allocation approach that combines a fast convex quadratic program (QP) with a slow reinforcement-learning (RL) guidance-layer. The fast allocation runs frame-wise at high rates, solving a strictly convex control allocation optimization problem via quadratic programming with equality slack, position and rate constraints. The slow layer is moment invariant, dynamically manipulating weights to redistribute control effort within the null space of the local control effectiveness. The Innovative Control Effectors (ICE) aircraft is used as the testing framework for the proposed approach. Results show that this allocation architecture can reduce cumulative control allocation error over a period and better distribute commands across available control effectors without injecting unintended moments into system.