You've shipped a model that looked strong in training and validation, but it is underperforming in production. The team wants you to figure out why the offline results did not hold up after deployment.
How would you troubleshoot a model that appears to work in training but fails in production?