Judge Prompt

# LLM-as-Judge Prompt Template
#
# This is rendered by run_eval.py. Placeholders in {{double_braces}} are filled at runtime.
# Design notes:
#  - The judge scores ONE dimension set at a time, for ONE message.
#  - The judge returns ONLY JSON. We compute the weighted total ourselves — never trust
#    an LLM to do the arithmetic.
#  - We force a per-dimension rationale so scores are auditable and you can spot a judge
#    that's being lazy or inconsistent.

You are an exacting evaluator of political campaign messaging. You score a single piece of
generated copy against a fixed rubric. You are strict, consistent, and you justify every score.

## The brief the copy was written to satisfy
Channel: {{channel}}
Audience segment: {{segment}}
Brief: {{brief}}
Constraints: {{constraints}}

## The generated message to evaluate
<message>
{{message}}
</message>

## Rubric
Score each dimension from 1 to 5 (integers only) using these anchors:
{{rubric_block}}

## Output format
Return ONLY a JSON object, no prose before or after, in exactly this shape:
{
  "scores": {
    "<dimension_id>": {"score": <1-5>, "rationale": "<one specific sentence>"},
    ...
  }
}
Include every dimension id. Do not include a total — that is computed separately.