Grade with code before a grading model

How to do it

Check the schema, enum values, required facts, forbidden phrases, citations and length in code. They are fast, free and never vary. For tone or helpfulness, write a grading prompt with a fixed rubric and a yes or no answer, not a score from 1 to 10. Have two people label 100 answers, then measure how often the grading model agrees with them. Use a different model, or at least a different prompt, from the one being graded.

Example

export function gradeReturnAnswer(answer: ReturnAnswer, expect: Expectation): string[] {
  const failures: string[] = [];
  if (answer.eligible !== expect.eligible) failures.push(`eligible was ${answer.eligible}, expected ${expect.eligible}`);
  for (const fact of expect.mustMention ?? []) if (!answer.reason.toLowerCase().includes(fact)) failures.push(`missing "${fact}"`);
  if (/\bguarantee(d)?\b/i.test(answer.reason)) failures.push('promises a guarantee');
  return failures;
}

def grade_return_answer(answer: ReturnAnswer, expect: dict) -> list[str]:
    failures = []
    if answer.eligible != expect["eligible"]:
        failures.append(f"eligible was {answer.eligible}, expected {expect['eligible']}")
    failures += [f'missing "{fact}"' for fact in expect.get("must_mention", []) if fact not in answer.reason.lower()]
    if re.search(r"\bguarantee(d)?\b", answer.reason, re.I):
        failures.append("promises a guarantee")
    return failures

How to check

Each eval case has at least one code check. The grading model's agreement with the human labels is stored and is at least 0.8.