🎯 场景：诊断 LLM judge 自身有没有偏见——长 vs 短、A 位置 vs B 位置、bullet vs prose、verbose vs terse 等。给一批 judge 输出，本卡返回每种偏见的检测结果 + 严重度。LLM-as-judge pipeline 上线前必备 sanity check。

Quick Use

Use when: You're using an LLM as a judge in production and want to verify it doesn't have systemic bias on length / position / format dimensions before trusting its scores. Fill in: {{judge_outputs}} = JSON array of judgment records (prompt, response, verdict, metadata fields). You'll get: Per-bias detection results (which biases detected, evidence, severity) and an overall trust signal. Output is JSON.

Purpose

Probe an LLM-as-judge for systematic biases that would invalidate its scores. Common biases tested: (1) length bias — does the judge prefer longer responses regardless of substance? (2) position bias — in pairwise judging, does the first-shown response win disproportionately? (3) format bias — does the judge prefer bullets / specific Markdown? (4) verbosity bias — does the judge reward "showing your work" over concise correctness? (5) self- preference — when judging outputs from the same model family.

Used as pre-launch sanity check on any LLM-as-judge pipeline. Output is structured so individual biases can be tracked over time and mitigation strategies (length normalization, position randomization) can be tied to specific biases detected.

Prompt

You diagnose biases in an LLM judge by analyzing a batch of its
judgment records. The data is already collected; you analyze it for
systematic patterns.

Judge outputs (batch of records, each with prompt / response /
verdict / metadata):
{{judge_outputs}}

Bias categories to check:
1. "length_bias"     : Does the judge favor longer responses?
                       Compare verdict distribution across response
                       length quartiles.
2. "position_bias"   : For pairwise records, does the judge prefer
                       the first-shown option disproportionately?
                       Need metadata.position field.
3. "format_bias"     : Does the judge prefer bulleted / formatted
                       responses over equivalent prose? Need
                       metadata.format_class field.
4. "verbosity_bias"  : Does the judge prefer "show your reasoning"
                       responses over concise correct ones? Look at
                       responses with reasoning_visible vs not.
5. "self_preference" : Does the judge prefer outputs from a specific
                       model family disproportionately? Need
                       metadata.model_family field.

For each bias category, classify detection:
- "not_detected"  : Distribution looks balanced, or insufficient
                    metadata to test.
- "weak"          : Slight skew (e.g. 5-10% above neutral baseline).
- "moderate"      : Clear skew (10-25% above baseline).
- "strong"        : Severe skew (>25% above baseline) — judge
                    scores are unreliable on this axis.

Provide concrete evidence: cite specific count comparisons or rate
gaps from the data.

Return ONLY this JSON object:
{
  "bias_results": [
    {
      "category": "<one of the 5>",
      "detection": "not_detected" | "weak" | "moderate" | "strong",
      "evidence": "<concrete count / rate observation from the data, e.g. 'top length quartile won 73% of pairwise vs bottom quartile 27%'>",
      "metadata_required": "<which metadata field needed; if missing in data, say 'metadata field X not present in inputs'>",
      "mitigation": "<for weak+ detections: short suggested mitigation>"
    }
  ],
  "overall_judge_trustworthy": true | false,
  "trustworthy_reason": "<one sentence>",
  "n_records_analyzed": <integer>,
  "decision_basis": "<one sentence, <=40 words, no internal CoT>"
}

Trust rule:
- overall_judge_trustworthy: false if any "strong" detected, OR
  two+ "moderate" detected.
- overall_judge_trustworthy: true otherwise.

Example

Input:

judge_outputs: [
  /* 50 pairwise records with metadata: position (a or b), response_a_length, response_b_length, judge picked: a or b */
  {"id": 1, "judge_picked": "a", "metadata": {"position_a_length": 320, "position_b_length": 180}},
  {"id": 2, "judge_picked": "a", "metadata": {"position_a_length": 50, "position_b_length": 480}},
  /* ... 48 more records ... */
]

Expected output:

{
  "bias_results": [
    {
      "category": "length_bias",
      "detection": "moderate",
      "evidence": "Across 50 pairwise records, the longer response won 32 / 50 (64%); for an unbiased judge expected ~25 / 50 (50%). Skew is 14 percentage points above neutral.",
      "metadata_required": "response_a_length and response_b_length present.",
      "mitigation": "Add 'do not reward verbosity' instruction; consider length-normalizing the prompt or pairing length-matched candidates."
    },
    {
      "category": "position_bias",
      "detection": "weak",
      "evidence": "First-position (A) won 28 / 50 (56%); marginal skew above 50% baseline.",
      "metadata_required": "position metadata present (always 'a' first).",
      "mitigation": "Run with positions swapped on a sample to confirm; if confirmed, randomize position at calling time."
    },
    {
      "category": "format_bias",
      "detection": "not_detected",
      "evidence": "metadata field 'format_class' not present in inputs; cannot test.",
      "metadata_required": "metadata.format_class (e.g. 'bulleted' / 'prose')",
      "mitigation": ""
    },
    {
      "category": "verbosity_bias",
      "detection": "not_detected",
      "evidence": "metadata field 'reasoning_visible' not present in inputs; cannot test directly. Length bias above suggests this may also be at play.",
      "metadata_required": "metadata.reasoning_visible (true/false)",
      "mitigation": ""
    },
    {
      "category": "self_preference",
      "detection": "not_detected",
      "evidence": "metadata field 'model_family' not present in inputs; cannot test.",
      "metadata_required": "metadata.model_family (e.g. 'gpt' / 'claude' / 'gemini')",
      "mitigation": ""
    }
  ],
  "overall_judge_trustworthy": false,
  "trustworthy_reason": "One moderate (length) and one weak (position) bias detected; combined effect makes pairwise scores unreliable until length is controlled.",
  "n_records_analyzed": 50,
  "decision_basis": "Length bias 64% (vs 50% neutral) is moderate; position bias 56% is weak; together unreliable for ranking decisions."
}

Failure Modes

• Insufficient data inferred as no bias — model says "not_detected" on small samples (<30) when bias might exist. Mitigation: track n_records_analyzed; on small samples, "not_detected" should be paired with a note about statistical power.
• Spurious detection on random noise — small sample with random outcome flagged as bias. Real bias detection needs ~100+ records for confidence; flag detections on small samples as preliminary.
• Metadata missing not flagged — model invents bias detection for categories where data lacks the required metadata. metadata_required field exists for honesty about this.
• Mitigation hand-waving — "improve the prompt" as mitigation. Reject vague mitigations; each should be a concrete action (length normalization, position randomization, prompt change).
• Aggregation method opaqueness — evidence says "the judge is biased" without citing concrete counts. Reject any evidence not containing numbers from the data.
• Trust signal flipping — overall_judge_trustworthy: true while multiple moderates detected. Verify rule logic at parse time.

Tuning Notes

• 模型差异：分析需要的统计推理能力中等；本卡更挑战的是"诚实地说出 data 中真实呈现的 pattern"。frontier 模型必须的，否则容易给出 好看但不准确的诊断。
• 温度：0.0，分析必须可重现。
• batch 规模建议：≥100 records per bias category for confidence。 <50 视为先导观察，需要再跑一批。
• metadata field 规划：上线前 instrument judge pipeline，捕获 position / response length / model_family / format_class / reasoning_visible 等字段。否则本卡能查的偏见有限。
• 与 eval/pairwise-judge-with-position-bias-probe 的关系：那张卡 在单次比较时主动控制 position bias（双向调用）；本卡在事后 分析一批 judgment 检测哪些偏见已经发生。前者预防，后者诊断。
• 与 eval/safety-output-classifier 的关系：safety 分类是输出层 评估；本卡是 judge 层评估。LLM-as-judge pipeline 一层比一层 meta， 审 judge 是质量控制的高阶环节。
• mitigation 落地：每种 bias 都有标准 mitigation。length-bias 用 length-normalization (length-controlled subset 重新跑判); position-bias 用 random shuffle 输入；format-bias 用 normalized format prompt。
• 上线前 baseline：保留一组人工 gold 标注的样本作 calibration reference。本卡跑判结果时连同 ground-truth 一起跑，看 judge 是否 与人类一致——单纯有 bias 不一定坏，关键看是否影响最终 metric。

Changelog

• 0.1.0 — Initial card.