A company has a curated dataset of input-output demonstrations for a specialized writing style and stable task format. It wants to adapt a supported model behavior across many requests, rather than retrieve frequently changing facts. Which customization approach most directly uses these demonstrations to update model behavior?

1. Supervised fine-tuning of a supported model using the curated demonstrations.
2. A vector index alone, with no model request or behavioral customization.
3. Increasing the output-token limit without supplying any examples.
4. Retraining a new broad foundation model from scratch as the only possible adaptation method.