A lightweight modification applied only during model inference without changing the underlying model weights.