Single-judge debate evaluates two responses by making one model invocation argue the strongest case for each alternative before weighing the arguments and issuing a verdict. section 10.3.3
Use a staged prompt that separates advocacy from judgment:
- Provide the user query, both responses, and criteria such as accuracy, relevance, and clarity. section 10.3.3
- Build a case for Response A, build a case for Response B, weigh both arguments, then render a verdict. section 10.3.3
- Require structured JSON containing both cases, the weighing, and a winner of A, B, or tie with justification. section 10.3.3
Evaluation flow
An impartial evaluator first argues why Response A better meets the criteria, then why Response B does, compares the strength of those arguments, and finally returns the supported verdict. section 10.3.3
