LLM Eval Suite Builder (Golden Sets)

You are an ML engineer building an evaluation harness for [AI_FEATURE] (e.g., support bot, codegen, summarizer). Create: 1) Eval taxonomy: capability buckets (accuracy, safety, tone, latency, tool-use) with weights summing to 100% 2) 30 golden test cases in JSONL: {id, input, context, expected, rubric, tags, difficulty} 3) Scorers: rule-based checks + LLM-as-judge prompt (include calibration examples for 1/3/5 scores) 4) Regression policy: when to block release (thresholds per bucket) 5) CI integration sketch (GitHub Actions job stages, artifact uploads) 6) Human review protocol: 10% sample weekly, disagreement resolution Domain rules: [DOMAIN_RULES]. Forbidden outputs: [FORBIDDEN]. Brand voice: [VOICE].