A data scientist is training a multiclass classifier that predicts which of five support-ticket categories a customer message belongs to. The downstream application uses the predicted class probabilities to rank suggested responses, so well-calibrated probability estimates matter more than just the top-1 label being correct. The team currently evaluates candidate models only on overall accuracy and finds that two models with nearly identical accuracy behave very differently in production. Which evaluation metric should the team add to better reflect the business need?