A model's resistance to being misled by incorrect external signals while still benefiting from correct or relevant context.