A demonstration combines GEPA, prompt changes, and labeled examples to adapt a model to specific decision criteria.
A demonstration described in the publication adapts the Jev model, trained on synthetic data, to specific decision criteria. The approach combines GEPA, prompt changes, and labeling a small number of examples.
To study the method, start by writing the decision criteria explicitly. Then gather representative examples and label them against those criteria; ambiguous examples can reveal where the rules need clarification.
Compare model responses before and after changes using the same examples. Record the prompt changes and assess whether results follow the criteria, rather than assuming that a few labeled examples guarantee consistent performance.
If you use an AI tool to analyze or apply this approach, avoid putting personal data or internal information in prompts and examples. Use synthetic or properly authorized data, and check your organization's rules before sharing content.