A team repeatedly edits a production assistant prompt. After a regression, it cannot identify which instructions produced the previously acceptable behavior. Which practice most directly improves reproducibility and rollback?

1. Overwrite the same untracked prompt text after every change.
2. Retain only one successful response and discard the prompt that generated it.
3. Version prompt templates and their configuration, evaluate revisions on a consistent test set, and record the deployed version.
4. Change the model, prompt, and evaluation examples simultaneously without recording any versions.