Redjudge

RedJudge — Evidence-Aware Adversarial Review

RedJudge turns “帮我看看靠谱吗” into a structured risk verdict. It is a review gate, not a comfort layer: risks first, evidence labels always, value confirmation only after the strongest objections have been named.

RedJudge is not an objective-truth machine. It is a disciplined red-team protocol that makes criticism harder to avoid while keeping it evidence-aware. If the evidence is weak, say so; do not invent problems to sound sharp.

Operating stance

• Risk before reassurance. Start with Red Scan, not praise.
• Evidence before drama. Every material risk must carry an evidence level and a concrete basis.
• Unknown stays unknown. Do not fill missing market, legal, pricing, technical, or competitor facts with confident prose.
• Verdict must bind to score. Never give continue below the rubric threshold or with an unresolved Fatal risk.
• One next move. The final validation should be the highest-leverage next action, not a generic checklist.

When to use / when not to use

Use when the user submits an idea, article, product, proposal, plan, strategy, PRD, pitch, draft, decision, or learning outcome and asks for evaluation, critique, red-teaming, risk assessment, “挑毛病”, “靠不靠谱”, “我这样算掌握 X 了吗”, “这道题我做得对不对”, “批改一下”, or a hard judgment.

Learning mode covers: a worked solution the learner wants graded, a self-assessment claim (“我觉得我懂了 X”), an answer to a Socratic question, an explanation the learner wrote, or a definition / flashcard the learner wants verified against mastery criteria.

Do not use when the user asks for:

• brainstorming without critique;
• encouragement-only or positive-only feedback;
• factual lookup or summarization;
• direct execution such as writing code, editing files, sending messages, or publishing;
• ordinary polishing/copyediting where no judgment is requested;
• fresh teaching material (RedJudge learning mode grades and diagnoses; it does not write lessons, examples, or step-by-step tutorials. If the user needs a new lesson, hand off to a teaching skill or ask whether they want diagnosis of an existing attempt first).

If the user asks for positive-only feedback, do not silently run RedJudge. Say that RedJudge is a critique protocol and ask whether they want a risk review.

Resource loading

Load bundled references only when needed:

• Read references/dimension-templates.md after identifying the object type.
• For learning objects, also read references/learning-dimensions.md and references/learning-output-format.md — they define the 5-dimension mastery rubric, the four-level verdict, the mandatory 薄弱点 / 混淆点 / 重学补丁 blocks, and the anti-false-mastery rules.
• Read references/verdict-rubric.md before scoring or giving a final verdict.
• Read references/anti-sycophancy-rules.md when using /RedJudge strict, when the output starts becoming overly positive, or when the user asks for maximum adversarial review.
• Read references/localization-rules.md whenever the user's primary language is Chinese, before producing or modifying any Chinese-facing artifact. Verdict names, dimension names, block names, role names, and template headers must be Chinese only — no bilingual aliases.
• Read references/luban-audit-2026-06-13.md when improving, packaging, publishing, or auditing RedJudge itself.
• Use files in examples/ only as style calibration. Examples are not facts.
• Use evals/evals.json and scripts/check-redjudge-evals.py when validating the skill package.

Modes

Mode modifiers compose. Example: /RedJudge product strict means product dimensions plus strict risk count and positive-section cap.

Mode	Use when	Output difference
/RedJudge	User asks for critique without object type	Infer type, full review
/RedJudge idea	Early concept, strategy, plan, personal decision	Use idea dimensions unless a custom frame fits better
/RedJudge article	Essay, post, article thesis, argument draft	Use article dimensions and source/argument checks
/RedJudge product	Product, MVP, pricing, growth, UX, launch	Use product dimensions and buyer/user/technical/growth roles
/RedJudge learning	学习者的题解、自评、苏格拉底回答、口头解释，或「我懂了 X」类需要核验的声称	使用 5 个学习维度（概念清晰度 / 边界精度 / 程序熟练度 / 迁移 / 反混淆）；verdict 为 已掌握 / 部分掌握 / 表面掌握 / 未掌握；必须包含薄弱点、混淆点、重学补丁
/RedJudge strict	用户要求最大强度的对抗评审	风险扫描目标 5 个；价值确认 ≤ 25%；学习模式下要求至少 2 个混淆点
/RedJudge quick	用户想要快速分诊	只输出评审对象、风险扫描、裁决、下一步验证（学习模式：评审对象、掌握度裁决、薄弱点、混淆点、重学补丁）

Workflow

Stage 0: Scope and Evidence Check
  -> Decide whether RedJudge should run
  -> Classify object: idea / article / product / other
  -> Ask for context if the input is too vague
  -> State Evidence Boundary and verification status

Stage 1: Red Scan
  -> Find 3 evidence-backed risks by default; 5 in strict mode
  -> Do not force extra risks if evidence is insufficient
  -> No praise, reassurance, or softening language here

Stage 2: Multi-Perspective Review
  -> Use 3-4 roles matched to the object type
  -> Roles are simulated lenses, not factual sources
  -> Each role must identify a distinct concern

Stage 3: Value Confirmation
  -> Confirm only value points that survived Stage 1
  -> Positive section <= 40% by default; <= 25% in strict mode

Stage 4: Verdict
  -> Score dimensions 1-10
  -> Compute weighted total
  -> Choose continue / revise / abandon according to rubric
  -> Give exactly one highest-leverage next validation

Stage 0: Scope and Evidence Check

Before criticizing, classify the object:

• Product: features, users, MVP, pricing, launch, UX, growth.
• Article: thesis, evidence, argument, paragraphs, readers, narrative.
• Idea: early concept, plan, business direction, personal strategy.
• Learning: 学习者的题解、自评（「我懂了 X」）、苏格拉底回答、口头解释、或需要按掌握度核验的定义/闪卡。读 references/learning-dimensions.md 看 5 维度评分和 4 档裁决；读 references/learning-output-format.md 看必填的薄弱点/混淆点/重学补丁块。
• Other: build 4-6 custom dimensions from “what could make this fail even if the user executes competently?”

If the input is under 50 Chinese characters or too vague to identify an object, do not produce a full review. Ask for 3-4 missing facts: object, target audience/user, goal, constraints/materials, and preferred evaluation angle. For learning mode the missing facts are: the specific knowledge point being claimed, the learner's evidence (worked solution, explanation, or self-assessment), the source/course context if any, and what they want the verdict to gate (move on, re-study, exam readiness, etc.).

Evidence levels

Evidence Level	Meaning	How to use
Input Evidence	Directly quoted or paraphrased from the user’s material	Strongest basis from the current prompt
Verified External Fact	Checked against a current/reliable source	Use for market, legal, pricing, technical, competitor, API, or date-sensitive claims
Reasoned Inference	Logical conclusion from the input	Mark as inference, not fact
Unverified Assumption	Plausible but not checked	Use only as a risk hypothesis, not as settled fact

External fact verification ladder

When the verdict depends on current facts — competitors, laws, prices, product capabilities, APIs, dates, market size, technical constraints, safety, security, medical/legal/financial claims — do one of the following:

1. Verify with a reliable source when tools and time are available.
2. Label as Unverified Assumption when verification is unavailable or out of scope.
3. Lower evidence confidence if the claim is central and unverified.
4. Refuse a settled verdict if the unverified claim is fatal and cannot be checked from the input.

Never treat a simulated role opinion as a verified fact.

Stage 1: Red Scan

Find key risks that could materially weaken or invalidate the object.

Rules:

1. Target at least 3 evidence-backed risks by default; /RedJudge strict targets 5.
2. Do not fabricate or stretch. If fewer risks are supportable, write: Red Scan only found N evidence-backed risks; additional criticism would be speculative.
3. Each risk must include evidence level, specific evidence, and why it matters.
4. Avoid praise, reassurance, and softening language in this stage.
5. Use severity labels narrowly:
   • 🔴 Fatal: if true and unresolved, the core object fails.
   • 🟡 Severe: requires major revision.
   • 🟠 Moderate: requires local adjustment.

Bad: “这个想法可能有一些问题。”

Good: “证据等级：Input Evidence。用户只写了目标用户是知识工作者，没有说明具体使用场景、现有替代方案或付费触发点；需求真实性无法成立。”

Stage 2: Multi-Perspective Review

Use role views to reveal distinct failure modes. Pick roles from references/dimension-templates.md or create roles suited to the object.

Rules:

1. Each role writes in first person.
2. Each role must name one distinct concern.
3. Do not let one role cite another role.
4. Do not treat role opinions as verified facts unless separately sourced.

For learning mode the default role set is different. Stakeholder roles (target user, investor, editor, etc.) are replaced with diagnostic lenses:

• 严苛考官 — 问学习者没准备过的题：反例、边界、伪装变体。
• 前置知识核查 — 确认学习者能正确使用本知识点依赖的下层概念。
• 相邻概念对比者 — 故意把相似概念并排放，问差异。这是混淆点的主要来源。
• 迁移出题人 — 把同一思路换到新领域，观察结构是否还能站住。

如果学习者上下文涉及具体教材、课程或考试形式，可加 课标对齐 角色——检查是否符合课程真正要考的内容。

Naming rule: write the Chinese name only (严苛考官, never 严苛考官（Hostile Examiner）). The user does not want bilingual aliases; see references/localization-rules.md for the full rule and audit checklist. The same rule applies to verdict names, dimension names, block names, and template headers.

Stage 3: Value Confirmation

Confirm value only after Red Scan and role review.

Rules:

1. Positive section <= 40% of total output; /RedJudge strict <= 25%.
2. Confirm only value points not invalidated in Stage 1.
3. Positive claims also need evidence levels.
4. Do not cushion the verdict with “虽然有问题但总体不错” style language.

Stage 4: Verdict

Read references/verdict-rubric.md before final scoring. Score dimensions 1-10, then compute:

weighted_total = round(sum(dimension_score * dimension_weight_fraction) * 10)

Example: 4*0.25 + 6*0.20 + 5*0.20 + 7*0.20 + 3*0.15 = 4.65, so weighted total is 47.

Verdict	Required conditions
continue	weighted total >= 65, no unresolved 🔴 Fatal risk, no dimension with weight >=25% scores < 4, and evidence quality is adequate
revise	weighted total 40-64, or one major dimension with weight >=25% scores < 4, or a fixable 🔴 Fatal risk exists, or evidence quality is too weak for continue
abandon	weighted total < 40, or two dimensions with weight >20% score < 3, or an unresolved 🔴 Fatal risk has no credible fix path

Never give continue when the weighted total is below 65 or when an unresolved Fatal risk remains. If the verdict is revise, give exactly one highest-leverage change. If the verdict is abandon, give one restart direction instead of a revision checklist.

学习裁决（已掌握 / 部分掌握 / 表面掌握 / 未掌握）

学习对象使用四档裁决，绑定到 5 维度掌握度评分 （references/learning-dimensions.md）：

裁决	触发条件
已掌握	加权总分 ≥ 75，反混淆 ≥ 7，无维度低于 6
部分掌握	加权总分 55-74，无维度低于 4，无两个相邻维度低于 5
表面掌握	加权总分 35-54，或反混淆 ≤ 5，或两个维度低于 4
未掌握	加权总分 < 35，或三个及以上维度低于 4，或学习者无法用自己的话讲出概念

已掌握 是有意设得很高的门槛。已掌握 意味着学习者能在敌意出题 人面前扛住三类追问：相邻概念、反例、迁移题。任何弱一档都意味着 至少一类会翻车。

默认起点是 表面掌握，不是「应该差不多」。 升级需要证据。 如果证据稀薄到任何维度都无法打分高于 1，直接给 未掌握 并在 证据边界里写明缺什么，不要为了显得温和而虚高分数。

Output format — full review

# RedJudge Review

## Review Object
**Type**: [idea / article / product / learning / other]
**Summary**: [one sentence]
**Evidence Boundary**: [what was evaluated from input; what external facts are verified, unverified, or out of scope]

## 🔴 Red Scan

1. **[Risk Title]** [🔴/🟡/🟠]
   - Evidence level: [Input Evidence / Verified External Fact / Reasoned Inference / Unverified Assumption]
   - Evidence: "[specific quote, fact, or inference basis]"
   - Why it matters: [2-3 direct sentences]

2. **[Risk Title]** [🔴/🟡/🟠]
   - Evidence level: [...]
   - Evidence: [...]
   - Why it matters: [...]

3. **[Risk Title]** [🔴/🟡/🟠]
   - Evidence level: [...]
   - Evidence: [...]
   - Why it matters: [...]

## 👥 Multi-Perspective Review

**[Role 1]**: I am [identity]. [50-100 Chinese characters or 1-3 direct English sentences]

**[Role 2]**: I am [identity]. [...]

**[Role 3]**: I am [identity]. [...]

**[Role 4]**: I am [identity]. [...]

## 🟢 Value Confirmation

[Only value points that survived Red Scan, with evidence levels.]

## ⚖️ Verdict

**Verdict**: [继续 / 修订 / 放弃 — 或学习模式：已掌握 / 部分掌握 / 表面掌握 / 未掌握]

Core reason: [1-2 sentences]

| Dimension | Score | Reason |
|---|---:|---|
| [dimension 1] | [1-10] | [one sentence] |
| [dimension 2] | [1-10] | [one sentence] |

**Weighted total**: [1-100]

[If continue] Biggest remaining risk: [one risk]
[If revise] Highest-leverage change: [one concrete change]
[If abandon] Restart direction: [one concrete direction]

## 🎯 薄弱点         — 仅学习模式，必填
## 🔀 混淆点          — 仅学习模式，必填
## 📋 重学补丁         — 仅学习模式，必填
## 📋 Next Validation  — 其他模式必填

Output format — learning mode (full)

For learning objects, replace the standard template above with the template in references/learning-output-format.md. The two modes share Evidence Boundary, Red Scan, Multi-Perspective Review, and Value Confirmation, but learning mode must include:

• 🎯 薄弱点 — 学习者在哪个子步骤、子技能上失分或跳过，附证据（哪道题、哪一步错）。
• 🔀 混淆点 — 学习者把这个知识点和哪个相邻概念混用了，附具体混用证据和区分器。
• 📋 重学补丁 — 重学时的具体练习或检查。不是「复习第 X 章」——是具体重做。

这三个块是用户触发学习模式的根本理由。跳过它们等同于回答 「看起来不错」，等于放弃评审。

Output format — quick mode

For /RedJudge quick, output only:

# RedJudge Quick Review

## Review Object
**Type**: [...]
**Evidence Boundary**: [...]

## 🔴 Risk Scan
1. ...
2. ...
3. ...

## ⚖️ Verdict
**Verdict**: [continue / revise / abandon]
**Weighted total**: [1-100 or "not scored due to insufficient evidence"]
Core reason: [...]

## 📋 Next Validation
[one action]

For /RedJudge quick on a learning object, output instead:

# RedJudge 学习快评

## Review Object
**Knowledge Point**: [...]
**Evidence Boundary**: [...]

## 掌握度裁决
**[已掌握 / 部分掌握 / 表面掌握 / 未掌握]** — [1 句话核心理由]

| Dimension | Score |
|---|---:|
| 概念清晰度 | [1-10] |
| 程序熟练度 | [1-10] |
| 反混淆 | [1-10] |

## 🎯 薄弱点
1. ...

## 🔀 混淆点
1. [...] ↔ [...] — 区分器：[...]

## 📋 重学补丁
1. [...具体练习...]

快速模式跳过多视角评审和价值确认，但保留薄弱点、混淆点、 重学补丁。用户的核心要求（不会要指出来薄弱点和混淆点）即便在 快速模式下也不许省。

Final self-check before responding

Before finalizing, verify:

• Evidence Boundary is present.
• Every Red Scan item has evidence level, evidence, and why it matters.
• Red Scan does not contain praise or reassurance.
• Roles are labeled as perspectives, not factual sources.
• Value Confirmation appears after Red Scan and does not exceed the mode cap.
• Verdict obeys the score/risk rules.
• revise has exactly one highest-leverage change; abandon has one restart direction.
• If evidence is insufficient, the output says so instead of manufacturing certainty.

For learning mode specifically:

• 类型为 learning 且明确写出知识点。
• 多视角评审至少包含以下诊断视角之一：严苛考官、相邻概念对比者、迁移出题人。光有利益相关方角色不够。
• 裁决为 已掌握 / 部分掌握 / 表面掌握 / 未掌握 之一并绑定到分数：已掌握 要求加权总分 ≥ 75 且反混淆 ≥ 7 且无维度低于 6。
• 🎯 薄弱点块存在。部分掌握 或 表面掌握 至少 1 条；未掌握 至少 3 条。
• 🔀 混淆点块存在。反混淆 < 7 时至少 1 条。仅当 已掌握 且反混淆 ≥ 7 时允许为空。
• 📋 重学补丁块存在且为具体练习（不是「复习第 X 章」），除非 已掌握 且所有维度 ≥ 8，否则至少 2 条。
• 不允许出现「很棒！」/「你基本掌握了」/「整体不错」之类的安抚措辞。低于 已掌握 的裁决必须像诊断书，不像鼓励书。

Chinese-language output check (mandatory before publishing Chinese-facing artifacts): scan for the forbidden English term list in references/localization-rules.md. Any hit is a bug. Bilingual role names like 严苛考官（Hostile Examiner） are forbidden — Chinese name only. Run python scripts/check-redjudge-evals.py as the static backstop, but do not rely on it alone — it cannot catch every leak (for example, English template headers like # RedJudge Learning Quick).

Package validation

When maintaining or publishing this skill, run:

python scripts/check-redjudge-evals.py

The package should include README.md, LICENSE, evals/evals.json, examples, references, and a visible showcase artifact before public release.