What would you do if you encountered a significant drop in model performance post-deployment?