What would you do if you discovered that a deployed model was underperforming in production?