global-reviewer

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Global Reviewer (survey draft)

Global Reviewer（调查草稿审查工具）

Purpose: make the draft read like a coherent paper (not stitched subsections) and make problems auditable.

目的：让草稿读起来像一篇连贯的论文（而非拼接的小节），并使问题可被审计。

Role cards (use explicitly)

角色卡片（需明确使用）

Consistency Reviewer (auditor)

Consistency Reviewer（审核员）

Mission: find cross-section issues a real reviewer would flag, and route them to the right upstream fix.

Do:

Check scope/taxonomy consistency and terminology drift across chapters.
Flag underspecified claims (numbers without task/metric/constraint).
Treat missing evidence as a failure signal; route upstream.

Avoid:

Writing around gaps by adding new claims or citations.

任务：找出真实审稿人会指出的跨章节问题，并将其导向正确的上游修复环节。

需执行：

检查各章节间的范围/分类一致性及术语偏差。
标记表述模糊的主张（无任务/指标/约束的数字）。
将缺失证据视为错误信号；导向上游处理。

需避免：

通过添加新主张或引用内容来掩盖漏洞。

Coherence Editor (bridge finder)

Coherence Editor（衔接编辑）

Mission: spot stitched-island structure and front-matter weaknesses that cause it.

Do:

Identify where transitions/leads are doing planner talk instead of argument bridges.
Flag repeated evidence-policy disclaimers and point to front matter as the single home.

Avoid:

Style-only nitpicks that do not change readability or verifiability.

任务：识别拼接式结构及导致该问题的前置内容缺陷。

需执行：

找出过渡/引导内容采用规划者话术而非论证衔接的地方。
标记重复出现的证据-政策免责声明，并指出前置内容为其唯一放置位置。

需避免：

仅涉及风格的细微修改，此类修改不影响可读性或可验证性。

Role prompt: Consistency Reviewer (AI paper reviewer mindset)

角色提示：Consistency Reviewer（AI论文审稿人思维）

text

You are a meticulous reviewer for a survey manuscript.

Your job is to surface cross-section problems that would matter to a real reader/reviewer:
- missing or underspecified evidence for claims
- scope drift and taxonomy inconsistency
- weak front matter (boundary/methodology not stated, so H3s carry repeated disclaimers)
- stitched-island structure (no argument chain across sections)

Constraints:
- do not invent facts or citations
- do not add new citation keys
- treat missing evidence as a failure signal: route upstream instead of writing around it

Output style:
- bullets-first
- actionable, route-to-skill recommendations

This is not “polish for style”. It is a contract check:

do claims align to evidence/citations?
do sections connect via a consistent lens?
does the front matter set the boundary and methodology so H3s can stay content-focused?

text

You are a meticulous reviewer for a survey manuscript.

Your job is to surface cross-section problems that would matter to a real reader/reviewer:
- missing or underspecified evidence for claims
- scope drift and taxonomy inconsistency
- weak front matter (boundary/methodology not stated, so H3s carry repeated disclaimers)
- stitched-island structure (no argument chain across sections)

Constraints:
- do not invent facts or citations
- do not add new citation keys
- treat missing evidence as a failure signal: route upstream instead of writing around it

Output style:
- bullets-first
- actionable, route-to-skill recommendations

这不是“风格润色”，而是合规性检查：

主张是否与证据/引用内容一致？
各章节是否通过统一视角衔接？
前置内容是否明确了边界和研究方法，从而让三级标题（H3）可以聚焦于内容本身？

Inputs

输入内容

```
output/DRAFT.md
```

Context (read-only; used to avoid drift):

```
outline/outline.yml
```
```
outline/taxonomy.yml
```
```
outline/mapping.tsv
```
```
outline/claim_evidence_matrix.md
```
```
citations/ref.bib
```

```
output/DRAFT.md
```

上下文（只读；用于避免偏差）：

```
outline/outline.yml
```
```
outline/taxonomy.yml
```
```
outline/mapping.tsv
```
```
outline/claim_evidence_matrix.md
```
```
citations/ref.bib
```

Outputs

输出内容

```
output/GLOBAL_REVIEW.md
```
(bullets-first report; always written)
```
output/DRAFT.md
```
(optional safe edits; only when edits are low-risk)

```
output/GLOBAL_REVIEW.md
```
（必生成，采用 bullet 点优先格式的报告）
```
output/DRAFT.md
```
（可选，仅进行低风险的安全编辑）

Non-negotiables

不可协商规则

No invented facts.
No invented citations.
Do not add/remove citation keys.
Missing evidence is a failure signal: write TODOs and route upstream; do not “write around” gaps.

不得编造事实。
不得编造引用内容。
不得添加/删除引用标识（citation keys）。
缺失证据视为错误信号：编写待办事项（TODOs）并导向上游处理；不得“掩盖”漏洞。

What this skill owns (and what it does not)

本工具的职责范围（及非职责范围）

Owns:

Cross-section coherence (throughline, definitions, scope)
Paper voice integrity (remove planner/pipeline narration where safe)
Terminology consistency (canonical term + synonym policy)
Claim→evidence hygiene (underspecified numbers, weak citations)

Does not own:

Changing the outline structure (route to C2)
Adding new sources/citations (route to C1/C4)
Strengthening missing evaluation details when notes are thin (route to C3/C4)

负责：

跨章节连贯性（主线、定义、范围）
论文语气一致性（在安全前提下移除规划者/流程式叙述）
术语一致性（规范术语+同义词规则）
主张→证据合规性（表述模糊的数字、薄弱引用）

不负责：

修改大纲结构（导向C2环节）
添加新来源/引用内容（导向C1/C4环节）
在笔记不足时补充缺失的评估细节（导向C3/C4环节）

Workflow (use the context files explicitly)

工作流程（需明确使用上下文文件）

Check structure against
```
outline/outline.yml
```

Verify the draft’s major sections and subsection order matches the intended ToC.
Identify which H2 is Introduction/Related Work so you can evaluate front-matter duties.

Check scope vocabulary against
```
outline/taxonomy.yml
```

Verify node descriptions and boundaries are consistent with how the draft uses the terms.
Flag mixed axes without a rule (model family vs capability vs evaluation).

Check coverage signals via
```
outline/mapping.tsv
```

Spot chapters/subsections that are under-mapped (likely under-cited or hollow).
Flag over-reuse of the same papers across many sections (suggests brittle synthesis).

Spot-check claims using
```
outline/claim_evidence_matrix.md
```

Sample 5–10 claims and verify each has plausible evidence fields and citations in the draft.
If the matrix is thin or mismatched, route upstream (C3/C4) instead of polishing prose.

Sanity-check citation keys against
```
citations/ref.bib
```

Flag undefined keys or suspicious naming (e.g., “GPT-5”) unless the cited work uses that label.

对照
```
outline/outline.yml
```
检查结构

验证草稿的主要章节及小节顺序是否与预期目录一致。
确定哪个二级标题（H2）是引言/相关工作，以便评估前置内容的职责。

对照
```
outline/taxonomy.yml
```
检查范围术语

验证节点描述和边界是否与草稿中的术语使用一致。
标记无规则的混合维度（如模型家族 vs 能力 vs 评估）。

通过
```
outline/mapping.tsv
```
检查覆盖信号

找出映射不足的章节/小节（可能引用不足或内容空洞）。
标记同一论文在多个章节中过度使用的情况（表明综合分析较为薄弱）。

利用
```
outline/claim_evidence_matrix.md
```
抽查主张

抽取5-10个主张，验证每个主张在草稿中是否有合理的证据字段和引用内容。
如果矩阵内容单薄或与草稿不匹配，导向上游环节（C3/C4）而非润色文稿。

对照
```
citations/ref.bib
```
检查引用标识

标记未定义的标识或可疑命名（如“GPT-5”），除非被引用的文献使用该标签。

Report format (required)

报告格式（必填）

output/GLOBAL_REVIEW.md

must be bullets-first and contain these headings verbatim (so gates can verify them):

## A. Input integrity / placeholder leakage

```
## B. Narrative and argument chain
```
```
## C. Scope and taxonomy consistency
```

## D. Citations and verifiability (claim -> evidence)

```
## E. Tables and structural outputs
```

Include a top line:

```
- Status: PASS
```
(or
```
- Status: OK
```
) only after all blocking issues are addressed.

output/GLOBAL_REVIEW.md

必须采用bullet点优先格式，并包含以下固定标题（以便关卡验证）：

```
## A. 输入完整性 / 占位符泄露
```
```
## B. 叙事与论证主线
```
```
## C. 范围与分类一致性
```

## D. 引用与可验证性（主张→证据）

```
## E. 表格与结构化输出
```

需包含顶部状态行：

```
- Status: PASS
```
（或
```
- Status: OK
```
）仅在所有阻塞性问题解决后使用。

What to check (high-value, paper-like)

检查要点（高价值，符合论文规范）

A. Input integrity / placeholder leakage

A. 输入完整性 / 占位符泄露

Look for:

leaked scaffolds (
```
…
```
,
```
TODO
```
, “enumerate 2-4 …”, “scope/design space/evaluation practice”)
planner talk in transitions or section openers
repeated evidence-policy boilerplate inside H3s

Action:

If placeholders exist: block and route upstream (do not patch them with “generic prose”).
If evidence-policy disclaimer repeats across H3s: move/keep it once in front matter and delete repeats.

检查内容：

暴露的脚手架内容（
```
…
```
、
```
TODO
```
、“enumerate 2-4 …”、“scope/design space/evaluation practice”）
过渡或小节开头的规划者话术
三级标题（H3）内重复出现的证据-政策模板内容

行动：

若存在占位符：阻塞流程并导向上游（不得用“通用文稿”填补）。
若证据-政策免责声明在多个H3中重复：将其移至前置内容中仅保留一份，并删除重复内容。

B. Narrative and argument chain

B. 叙事与论证主线

Goal: every section does an argument move.

Check:

H2 throughline: Introduction defines the boundary and evaluation lens; chapters execute comparisons; Discussion synthesizes cross-cutting risks/gaps.
H3 “argument shape”: tension → contrast → evaluation anchor → synthesis → limitation.
“Generator voice”: narration templates (
```
This subsection ...
```
) and slide navigation (
```
Next, we ...
```
).

Action (safe edits allowed):

Replace navigation sentences with argument bridges (no new facts).

Bad:

```
Next, we move from planning to memory.
```

Better:

Planning specifies how decisions are made; memory determines what information those decisions can reliably condition on under a fixed protocol.

目标：每个章节都承担论证环节的作用。

检查内容：

二级标题（H2）主线：引言定义边界和评估视角；各章节进行对比分析；讨论部分综合跨领域风险/漏洞。
三级标题（H3）的“论证结构”：矛盾→对比→评估锚点→综合→局限性。
“生成器语气”：叙述模板（
```
This subsection ...
```
）和幻灯片式导航（
```
Next, we ...
```
）。

允许的安全行动：

用论证衔接句替代导航句（不得添加新事实）。

反面示例：

```
Next, we move from planning to memory.
```

优化示例：

Planning specifies how decisions are made; memory determines what information those decisions can reliably condition on under a fixed protocol.

C. Scope and taxonomy consistency

C. 范围与分类一致性

Check:

Scope boundary is explicit and consistent (what counts as an “agent” here; what does not).
Taxonomy nodes match the paper’s claims (no mixed axes without a rule).
No silent drift (e.g., includes lots of multi-agent safety papers when scope is tool-use agents).

Action:

If scope drift is structural: route to C2 (tighten outline + mapping).
If scope drift is minor: tighten one scope sentence in the front matter (no new citations).

检查内容：

范围边界是否明确且一致（此处哪些属于“agent”，哪些不属于）。
分类节点是否与论文主张匹配（无规则的混合维度）。
是否存在隐性偏差（例如，当范围为工具型agent时，却包含大量多agent安全相关论文）。

行动：

若范围偏差为结构性问题：导向C2环节（收紧大纲+映射）。
若范围偏差为轻微问题：收紧前置内容中的范围陈述（不得添加新引用）。

D. Citations and verifiability (claim -> evidence)

D. 引用与可验证性（主张→证据）

Write a small claim-evidence table (5–10 rows):

claim | section | citations | evidence_field | evidence_level

Flag:

cite dumps and paragraphs with weak/irrelevant citations
underspecified quantitative claims (numbers without task/metric/constraint context)
ambiguous model naming (e.g., “GPT-5”) unless the cited paper uses that label

Action:

If you can clarify context without new facts (e.g., “under a fixed budget/tool access”), do so.
Otherwise: mark as TODO and route to C3/C4 (paper notes / evidence packs).

编写小型主张-证据表格（5-10行）：

claim | section | citations | evidence_field | evidence_level

标记：

引用堆砌和引用薄弱/无关的段落
表述模糊的量化主张（无任务/指标/约束上下文的数字）
模糊的模型命名（如“GPT-5”），除非被引用的文献使用该标签

行动：

若无需添加新事实即可明确上下文（如“在固定预算/工具访问权限下”），可进行补充。
否则：标记为TODO并导向C3/C4环节（论文笔记/证据包）。

E. Tables and structural outputs

E. 表格与结构化输出

Check:

Tables answer a concrete comparison question (schema), not copied outline bullets.
Rows contain citations.

Action:

If tables are intermediate-only in this pipeline run: ensure the draft does not contain thin “table placeholder” chapters.

检查内容：

表格是否回答具体的对比问题（符合 schema），而非复制大纲bullet点。
表格行包含引用内容。

行动：

若表格仅为本次流程中的中间产物：确保草稿不包含内容单薄的“表格占位符”章节。

Recommended fix order (routing)

Safe edits allowed (optional)

允许的安全编辑（可选）

If and only if edits are low-risk and do not change citation keys:

unify terminology
remove slide-like narration and planner talk
add 1–2 short argument-bridging transitions between major sections
tighten scope statements and conclusion closure

仅在编辑为低风险且不修改引用标识（citation keys）的前提下：

统一术语
删除幻灯片式叙述和规划者话术
在主要章节间添加1-2个简短的论证衔接过渡句
收紧范围陈述和结论收尾

Script

脚本

This skill includes a deterministic helper script that generates a gate-compliant

output/GLOBAL_REVIEW.md

from the current draft and context (no invented facts/citations).

本工具包含一个确定性辅助脚本，可从当前草稿和上下文生成符合关卡要求的

output/GLOBAL_REVIEW.md

（不会编造事实/引用内容）。

Quick Start

快速开始

python .codex/skills/global-reviewer/scripts/run.py --help

python .codex/skills/global-reviewer/scripts/run.py --workspace workspaces/<ws>

python .codex/skills/global-reviewer/scripts/run.py --help

python .codex/skills/global-reviewer/scripts/run.py --workspace workspaces/<ws>

All Options

所有选项

```
--workspace <dir>
```
```
--unit-id <U###>
```
(optional; for logs)
```
--inputs <semicolon-separated>
```
(rare override; prefer defaults)

--outputs <semicolon-separated>

(rare override; default writes

output/GLOBAL_REVIEW.md

)

```
--checkpoint <C#>
```
(optional)

```
--workspace <dir>
```
```
--unit-id <U###>
```
（可选；用于日志）
```
--inputs <semicolon-separated>
```
（罕见覆盖；优先使用默认值）

--outputs <semicolon-separated>

（罕见覆盖；默认生成

output/GLOBAL_REVIEW.md

）

```
--checkpoint <C#>
```
（可选）

Examples

示例

Generate a global review after merging a draft:

python .codex/skills/global-reviewer/scripts/run.py --workspace workspaces/<ws>

Freeze policy:

If you hand-edit the review and want to freeze it, create
```
output/GLOBAL_REVIEW.refined.ok
```
to prevent overwrites.

Notes:

The script does not “write” new survey content; it summarizes integrity/citation/structure signals and re-runs draft quality checks.

合并草稿后生成全局审查报告：

python .codex/skills/global-reviewer/scripts/run.py --workspace workspaces/<ws>

冻结规则：

如果你手动编辑了审查报告并希望冻结它，请创建
```
output/GLOBAL_REVIEW.refined.ok
```
以防止被覆盖。

注意：

该脚本不会“撰写”新的调查内容；它会总结完整性/引用/结构信号，并重新运行草稿质量检查。

Troubleshooting

故障排除

Issue: review flags missing citations / undefined keys

问题：审查报告标记缺失引用 / 未定义标识

Fix:

Run

citation-verifier

and ensure

citations/ref.bib

contains every cited key in

output/DRAFT.md

修复：

运行

citation-verifier

，确保

citations/ref.bib

包含

output/DRAFT.md

中的所有引用标识。

Issue: review suggests changes that would add new claims

问题：审查报告建议添加新主张的修改

Fix:

Convert those into “missing evidence” TODOs instead; this pass must not invent facts or citations.

修复：

将这些建议转换为“缺失证据”的TODO项；本次检查不得编造事实或引用内容。

global-reviewer

Original

Translation

Global Reviewer (survey draft)

Global Reviewer（调查草稿审查工具）

Role cards (use explicitly)

角色卡片（需明确使用）

Consistency Reviewer (auditor)

Consistency Reviewer（审核员）

Coherence Editor (bridge finder)

Coherence Editor（衔接编辑）

Role prompt: Consistency Reviewer (AI paper reviewer mindset)

角色提示：Consistency Reviewer（AI论文审稿人思维）

Inputs

输入内容

Outputs

输出内容

Non-negotiables

不可协商规则

What this skill owns (and what it does not)

本工具的职责范围（及非职责范围）

Workflow (use the context files explicitly)

工作流程（需明确使用上下文文件）

Report format (required)

报告格式（必填）

What to check (high-value, paper-like)

检查要点（高价值，符合论文规范）

A. Input integrity / placeholder leakage

A. 输入完整性 / 占位符泄露

B. Narrative and argument chain

B. 叙事与论证主线

C. Scope and taxonomy consistency

C. 范围与分类一致性

D. Citations and verifiability (claim -> evidence)

D. 引用与可验证性（主张→证据）

E. Tables and structural outputs

E. 表格与结构化输出

Recommended fix order (routing)

推荐修复顺序（路由）

Safe edits allowed (optional)

允许的安全编辑（可选）

Script

脚本

Quick Start

快速开始

All Options

所有选项

Examples

示例

Troubleshooting

故障排除

Issue: review flags missing citations / undefined keys

问题：审查报告标记缺失引用 / 未定义标识

Issue: review suggests changes that would add new claims

问题：审查报告建议添加新主张的修改