A company needs a model to consistently produce its approved report style. Facts come from a fresh retrieval corpus, but few-shot prompting still misses the required style. Which next experiment is most appropriate?

1. Increase temperature without evaluating output format.
2. Evaluate supported fine-tuning using curated style examples while retaining retrieval for current facts.
3. Fine-tune on unreviewed production errors as if they were correct target outputs.
4. Replace the retrieval corpus with the style guide and assume all factual knowledge stays current.