# Judge Audit

> Performs independent verification, causal chain audit, analogy validation, and generates structured rulings for debate rounds. Use this skill when the judge agent needs to audit a debate round, verify claims independently, check causal validity of arguments, produce a structured ruling, identify mandatory response points for next round, perform independent second-pass verification, evaluate reasoning quality, assess historical wisdom references, or review speculative scenarios.

- Skill: `finger88/judge-audit` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add finger88/judge-audit`
- Raw SKILL.md: https://api.skillmd.com/api/skills/finger88/judge-audit/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: MIT-0
- Author: finger88 (https://skillmd.com/u/finger88)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/finger88/judge-audit

---


## Changelog / 变更日志

| 时间 / Time | 作者 / Author | 变更 / Change |
|---|---|---|
| 2026-03-11 | Claude | v0.6.0: Agent Skills open standard compliance — frontmatter restructured, English-only description, progressive disclosure, evals added / Agent Skills 开放标准兼容 — 前置元数据重构、纯英文描述、渐进式披露、添加评测 |
| 2026-03-11 | Claude | v0.5.0: recovered from broken symlink, unified version / 从断开的 symlink 恢复，统一版本号 |

# JudgeAudit
# 裁判审计

Independent verification, causal chain audit, analogy validation, and structured ruling for a debate round.
独立验证、因果链审计、类比验证和辩论回合结构化裁定。

## When to Use / 何时使用

- After both Pro and Con have completed their turns in a round / 正反方均完成其回合后
- This is the Judge's primary workflow / 这是裁判的主要工作流
- See `references/data-contracts.md` for JudgeRuling output schema

## Input / 输入

- `pro_turn_path`: Path to rounds/round_N/pro_turn.json
- `con_turn_path`: Path to rounds/round_N/con_turn.json
- `evidence_store_path`: Path to evidence_store.json
- `claim_ledger_path`: Path to claim_ledger.json
- `round_number`: Current round

## Output / 输出

- `JudgeRuling` JSON written to rounds/round_N/judge_ruling.json
- Contains: verification_results, causal_validity_flags, mandatory_response_points, historical_wisdom_assessment, round_summary

## Core Workflow / 核心工作流

### Step 1: Independent Source Verification (CRITICAL) / 独立来源验证（关键）

**Do NOT trust debaters' citations at face value / 不要轻信辩手的引用**

For each factual claim from BOTH sides:

1. Read the claim and its referenced evidence_ids
2. Fetch the original evidence items from evidence_store
3. **Independently verify**: Use WebSearch and WebFetch to check:
   - Does the cited source actually support this claim? (not misquoted/out-of-context)
   - Is the source still accessible and current?
   - Are there counter-sources the debater didn't mention?
4. Use the EvidenceVerify skill with `independent_search = true`
5. Use FreshnessCheck to verify timeliness of fact-track claims
6. Produce a `verification_result` entry with recommended status transition
6b. When independent verification reveals conflicting sources for a claim:
   - Document the conflict in the verification_result's `reasoning` field
   - Structure it as: "Source A (evi_xxx) states [position]. Source B (evi_yyy) states [position]. Divergence point: [specific disagreement]."
   - This structured reasoning will be extracted by ClaimLedgerUpdate into `conflict_details`

**Priority**: Focus verification effort on:
- Claims central to the argument (high impact)
- Claims with only one supporting source
- Claims using tier3/tier4 sources
- Claims where debater's reasoning seems to stretch the evidence

### Step 2: Causal Chain Audit / 因果链审计

For each argument's reasoning chain from both sides, evaluate:

1. **Observed facts**: Are they accurate? Are they cherry-picked?
2. **Mechanism**: Is the causal explanation logically sound?
   - Flag: correlation presented as causation
   - Flag: reverse causality
   - Flag: confounding variables ignored
   - Flag: mechanism relies on unstated assumptions
3. **Scenario implication**: Does it logically follow from the mechanism?
4. **Trigger conditions**: Are they specific and testable?
   - Flag: vague or unfalsifiable triggers
5. **Falsification conditions**: Are they genuinely falsifiable?
   - Flag: conditions that are impossible to test
   - Flag: conditions so broad they're meaningless

Produce `causal_validity_flags` for each issue found, with severity:
- `critical`: Fundamental logical flaw that invalidates the argument
- `moderate`: Weakness that significantly undermines the argument
- `minor`: Small issue that should be addressed but doesn't invalidate

### Step 3: Analogy Audit / 类比审计

Apply AnalogySafeguard rules to all historical/classical analogies:

1. Check structural requirements (≥2 similarities, ≥1 difference)
2. Assess content share proportion
3. Mark invalid analogies as "heuristic only" in the ruling

### Step 3.5: Historical Wisdom Assessment (v3) / 历史智慧评估

For each `historical_wisdom` reference from both sides:

1. **Structural relevance check**: Does the historical case genuinely parallel the current topic, or is it a superficial connection?
2. **Difference honesty check**: Are the `key_differences` genuinely acknowledged, or minimized to make the parallel seem stronger?
3. **Lesson validity check**: Does the extracted lesson logically follow from the historical case?
4. **Cross-side comparison**: If both sides cite conflicting historical lessons, note the tension

Produce `historical_wisdom_assessment` in JudgeRuling:
```json
{
  "historical_wisdom_assessment": [
    {
      "side": "pro | con",
      "historical_event": "...",
      "relevance_grade": "strong_parallel | moderate_parallel | weak_parallel",
      "honesty_grade": "honest | partially_honest | misleading",
      "note": "..."
    }
  ]
}
```

**Key principle / 关键原则:** Judge evaluates the QUALITY of historical reasoning, not whether the historical lesson supports pro or con. A well-analyzed weak parallel is better than a poorly analyzed strong one.
Judge 评估历史推理的质量，而非历史教训支持正方还是反方。

### Step 3.6: Speculative Scenarios Review (v3) / 推演场景审查

For each `speculative_scenario` from both sides:

1. **Internal consistency**: Does the chain of events make causal sense?
2. **Probability calibration**: Is the probability estimate reasonable given the premise?
3. **Falsifiability**: Is `what_would_falsify` genuinely testable?
4. **Novelty**: Does this scenario add new insight, or just restate existing arguments as "what if"?

Do NOT verify or score speculative scenarios — they are by definition unverifiable. Instead, note quality of reasoning.
不要验证或评分推演场景 — 它们本质上不可验证。仅评估推理质量。

### Step 4: Mandatory Response Check / 必答点回应检查

If this is round 2+, verify that both sides addressed the previous round's mandatory response points:

1. Read previous round's judge_ruling mandatory_response_points
2. Check each debater's `mandatory_responses` array
3. Note any unaddressed points (the orchestrator should have caught this, but double-check)

### Step 5: Generate Mandatory Response Points / 生成必答点

Identify 2-5 points that MUST be addressed in the next round:

For each point:
- `point_id`: `mrp_<round>_<sequence>` (e.g., `mrp_1_1`)
- `target`: "pro" | "con" | "both"
- `point`: Clear statement of what must be addressed
- `reason`: Why this requires a response

**Types of mandatory response points / 必答点类型**:
- Unresolved factual contradictions between the sides
- Claims with critical causal validity issues
- New evidence that contradicts a key argument
- Weakly supported claims that need strengthening or withdrawal
- Questions raised by one side that the other didn't address

### Step 6: Round Summary / 回合总结

Write a neutral, balanced summary of the round:

- Key arguments from each side
- Which claims were verified, contested, or remain unverified
- The most important developments in this round
- No scoring, no preference — evaluate REASONING QUALITY, not which conclusion is "right"

**Conclusion Profile Data Notes (v3) / 结论画像数据备注:**

For each claim evaluated in this round, note in verification_results:
- Whether the causal chain is clear or has gaps (feeds into `causal_clarity`)
- Whether the claim has strong falsification conditions (feeds into `falsifiability`)
- How much new evidence would be needed to reverse the status (feeds into `reversibility`)

These notes will be consumed by FinalSynthesis when building Conclusion Profiles.

### Step 7: Produce JudgeRuling JSON / 生成 JudgeRuling JSON

Assemble all outputs into the JudgeRuling schema and write to the designated path.
Validate with `scripts/validate-json.sh <file> judge_ruling`.

## Impartiality Rules (NON-NEGOTIABLE) / 公正规则（不可协商）

1. **NEVER** express preference for either side
2. Evaluate **reasoning quality**, not which conclusion you personally favor
3. Apply **identical standards** to both sides
4. If one side is clearly stronger this round, say so neutrally with evidence
5. **NEVER** use emotional language in the ruling
6. **NEVER** introduce your own arguments — only evaluate what was presented
7. Hold BOTH sides to the same evidence standards (tier requirements, cross-source, etc.)