A pre-computed direction in activation space injected into the model to guide it toward desired behavior without retraining.