A model's input contains new categories outside the training vocabulary and an increased missing-value rate. Which approach best fits?

1. Monitor schema and data-quality conditions alongside distribution drift.
2. Evaluate answer support and citation relevance, not only readability.
3. Evaluate quality, resource cost, and serving objectives together.
4. Define an accountable response process for model-quality incidents.