judge-evals

name	judge-evals
description	Use when creating evals and judge checks for API QA agent reports.

Judge and Evals Skill

The judge exists to prevent fake confidence.

Required report shape:

{
  "happy_path": {"status": "PASS|FAIL", "steps": []},
  "adversarial": [
    {"rule": "overpayment_protection", "status": "HELD|BREACHED|INCONCLUSIVE", "evidence": {}}
  ],
  "summary": "..."
}

Eval checklist:

• Report parses as JSON.
• Contains happy_path, adversarial, summary.
• Every step has request, response, status code, interpretation.
• Every stated financial invariant has at least one adversarial finding.
• Negative case HELD includes evidence of clean rejection and/or valid state after.
• 500 responses are not treated as clean HELD guardrails.
• Any BREACHED finding includes reproduction evidence.
• Critical mutations re-fetch state where endpoint exists.

Use scripts/eval_report.py report.json to run basic structural evals.