Reasoning Up the Instruction Ladder for Controllable Language Models