A web application must return a model prediction within the current user request. Which inference pattern most directly fits?

1. A suitably provisioned real-time serving endpoint
2. A model registry description without deployed compute
3. A manual export reviewed after the user session ends
4. A once-per-month offline training report