To adapt ML models to concept drift under strict resource constraints, we propose a lightweight drift-plus-penalty policy that provably limits resource usage and achieves robust results.