An endpoint's error rate rises after an artifact update. Which records are most useful for diagnosis?

1. Only the number of notebook cells
2. Deployment logs and metrics correlated with the model and environment version
3. Only the model's descriptive tags
4. Only the original training score