Composite scoring

Score independent dimensions and make the final weighting visible in code.

SCORE×3

One request, 3 parallel answers (three shown). The model never performs the side effect — your code reads the confidences and decides.

HOW IT WORKS

questions = {
    "technical_depth": Score(
        instructions="How strong is the candidate's evidence of the required technical depth?",
        criteria=["No evidence", "Some exposure", "Regular applied experience", "Deep and recent expertise"],
    ),
    "system_design": Score(
        instructions="How strong is the candidate's evidence of designing reliable systems?",
        criteria=["No evidence", "Some exposure", "Regular applied experience", "Deep and recent expertise"],
    ),
    "communication": Score(
        instructions="How clearly does the candidate communicate technical trade-offs?",
        criteria=["Unclear", "Mixed", "Clear", "Exceptionally clear"],
    ),
}


def senior_ic_score(response) -> float:
    technical = response.answers["technical_depth"].score / 3
    design = response.answers["system_design"].score / 3
    communication = response.answers["communication"].score / 3
    return 0.45 * technical + 0.40 * design + 0.15 * communication

Keep the raw dimensions alongside the composite. When the result is surprising, you want to know whether the problem was evidence, a question, a weight, or a threshold. For recruiting and other high-impact domains, use explicit job-related criteria, audit for bias, and keep a human decision-maker in the loop.

Source: Composite scoring.

요청 1 / 답 3

이 패턴을 한 줄로