5 things I check before trusting a machine learning model's accuracy: 1. Is the test set truly unseen? 2. Class imbalance handled? (I use SMOTE where needed) 3. Does ROC-AUC tell the full story, or do I need precision/recall too? 4. Does it generalize outside the training data's time window? 5. Can I explain the top features to a non-technical client?