paper-review

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Paper Review

论文评审

Overview

概述

Use this skill to evaluate a paper as a reviewer would: identify evidence-bound weaknesses, rank fixes by risk and cost, prepare grounded rebuttals, and verify that promised revisions are actually traceable. Keep this skill separate from

paper-writing

paper-figures-tables

, and any external literature-management workflow.

使用本技能以审稿人的视角评估论文：识别基于证据的缺陷，按风险和修复成本排序需修正的问题，准备有依据的反驳内容，并验证承诺的修订内容是否可追溯。请将本技能与

paper-writing

、

paper-figures-tables

以及任何外部文献管理工作流区分使用。

Boundaries

边界限制

Do not invoke external literature-management tools or separate review/rebuttal builders by default. Use them only when the user explicitly asks for that separate workflow.
Do not perform automatic citation retrieval or reference verification. Use the citation and evidence sources declared by the user or project; local
```
.bib
```
, manuscript text, and supplied notes are the default.
Do not invent experiments, numbers, p-values, baselines, reviewer comments, citations, links, or revision status.
Do not rewrite prose as the primary deliverable. Route substantial prose drafting to
```
paper-writing
```
.
Do not redraw or typeset figures/tables as the primary deliverable. Route artifact creation or revision to
```
paper-figures-tables
```
.
Do not submit reviews, rebuttals, or OpenReview/CMT/HotCRP responses for the user. Produce drafts and checks for user sign-off.
Do not treat local house policy as a rejection criterion when formally reviewing someone else's paper. Apply the policy readiness gate to the user's own manuscript workflows only.

默认请勿调用外部文献管理工具或独立的评审/反驳构建工具。仅当用户明确要求使用独立工作流时才使用它们。
请勿自动检索引文或验证参考文献。默认使用用户或项目声明的引文和证据来源：本地
```
.bib
```
文件、手稿文本和提供的笔记。
请勿编造实验、数据、p值、基准、审稿人意见、引文、链接或修订状态。
请勿将重写文稿作为主要交付成果。如需大量文稿撰写，请转至
```
paper-writing
```
技能处理。
请勿将重新绘制或排版图表作为主要交付成果。如需创建或修订相关制品，请转至
```
paper-figures-tables
```
技能处理。
请勿代用户提交评审意见、反驳内容或OpenReview/CMT/HotCRP回复。仅生成草稿和检查内容供用户确认。
在正式评审他人论文时，请勿将本地内部政策作为拒稿标准。仅在用户自己的手稿工作流中应用政策就绪检查。

Source Priority

来源优先级

When reviewing or drafting responses, use evidence in this order:

Core integrity constraints from the sibling
```
paper-policy
```
skill.
User-confirmed facts, experimental results, revision decisions, and author intent.
Verified venue or template requirements with source and freshness date.
The current manuscript, local LaTeX/PDF/source files, local tables/figures, and local
```
.bib
```
.
Reviewer comments and local rebuttal drafts.
Rules from explicitly activated policy sets and the skill references listed below.

If a likely needed citation is missing locally, use the manual handoff format in

references/citation-and-evidence-policy.md

. Do not search, fetch, or validate references unless the user explicitly asks for that separate task.

在评审或撰写回复时，按以下优先级使用证据：

关联技能
```
paper-policy
```
中的核心完整性约束。
用户确认的事实、实验结果、修订决策和作者意图。
已验证的会议/期刊或模板要求，需包含来源和更新日期。
当前手稿、本地LaTeX/PDF/源文件、本地图表和本地
```
.bib
```
文件。
审稿人意见和本地反驳草稿。
已明确激活的政策集以及下文列出的技能参考中的规则。

如果本地缺少可能需要的引文，请使用

references/citation-and-evidence-policy.md

中的手动交接格式。除非用户明确要求执行独立任务，否则请勿搜索、获取或验证参考文献。

Mode Routing

模式路由

Read

references/review-modes.md

for mode details, then load only the references needed for the request.

Pre-submission audit: read

policy-compliance.md

pre-submission-audit.md

quality-gates.md

, and

citation-and-evidence-policy.md

Red-team my draft / simulate reviewers: read
```
policy-compliance.md
```
,
```
reviewer-panel.md
```
,
```
issue-board.md
```
, and
```
quality-gates.md
```
; use policy as a diagnostic surface, not as proof of reviewer reaction.
Formal review of someone else's paper: read
```
reviewer-panel.md
```
,
```
pre-submission-audit.md
```
, and
```
quality-gates.md
```
; keep tone fair and evidence-bound.

Analyze external reviews: read

issue-board.md

rebuttal-strategy.md

, and

citation-and-evidence-policy.md

Draft or polish rebuttal: read

rebuttal-strategy.md

rebuttal-drafting.md

revision-plan.md

, and

quality-gates.md

Verify revised manuscript after rebuttal: read
```
policy-compliance.md
```
,
```
revision-plan.md
```
,
```
quality-gates.md
```
, and the relevant writing or figure/table handoff rules.

阅读

references/review-modes.md

了解模式详情，然后仅加载请求所需的参考内容。

投稿前审核：阅读

policy-compliance.md

、

pre-submission-audit.md

、

quality-gates.md

和

citation-and-evidence-policy.md

。

红队评审我的草稿/模拟审稿人：阅读
```
policy-compliance.md
```
、
```
reviewer-panel.md
```
、
```
issue-board.md
```
和
```
quality-gates.md
```
；将政策用作诊断依据，而非证明审稿人反应的证据。
正式评审他人论文：阅读
```
reviewer-panel.md
```
、
```
pre-submission-audit.md
```
和
```
quality-gates.md
```
；保持语气公正、基于证据。

分析外部评审意见：阅读

issue-board.md

、

rebuttal-strategy.md

和

citation-and-evidence-policy.md

。

撰写或润色反驳内容：阅读

rebuttal-strategy.md

、

rebuttal-drafting.md

、

revision-plan.md

和

quality-gates.md

。

反驳后验证修订手稿：阅读
```
policy-compliance.md
```
、
```
revision-plan.md
```
、
```
quality-gates.md
```
以及相关的写作或图表交接规则。

Default Workflow

默认工作流

Identify the mode, manuscript stage, venue if known, deadline pressure, and available inputs.
For the user's own audit or revision-verification workflow, resolve policy and run the read-only compliance assessment in
```
policy-compliance.md
```
.
Establish an evidence map: manuscript locations, local results, local
```
.bib
```
, reviewer comments, user-confirmed constraints.
Atomize failures and unresolved evidence into issue cards instead of free-form critique.
Classify each issue by severity, policy status, evidence anchor, affected claim, and required action.
Apply quality gates: provenance, coverage, commitment, tone, fairness, citation policy, readiness, and handoff routing.
Return the most actionable artifact for the mode: compliance audit, reviewer-panel report, issue board, rebuttal strategy, rebuttal draft, or revision checklist.

确定模式、手稿阶段、已知的投稿会议/期刊、截止日期压力和可用输入内容。
对于用户自己的审核或修订验证工作流，解析政策并运行
```
policy-compliance.md
```
中的只读合规性评估。
建立证据映射：手稿位置、本地结果、本地
```
.bib
```
文件、审稿人意见、用户确认的约束条件。
将问题和未解决的证据拆解为原子化的问题卡片，而非自由形式的评论。
按严重性、政策状态、证据锚点、受影响的主张和所需行动对每个问题进行分类。
应用质量检查：来源、覆盖范围、承诺、语气、公正性、引文政策、就绪情况和交接路由。
返回对应模式下最具可操作性的制品：合规性审核报告、审稿人小组报告、问题看板、反驳策略、反驳草稿或修订检查清单。

Required Review Discipline

必备评审准则

Every criticism must point to a manuscript location, reviewer quote, local artifact, or
```
MISSING
```
marker.
Keep hard results in PASS, FAIL, UNVERIFIED, NOT_APPLICABLE, or WAIVED. Never convert missing evidence into PASS.
Record an agent semantic/manual FAIL only when the manuscript provides a precise artifact, locator, and contradictory evidence. Never use an agent record to assign PASS, WAIVED, or NOT_APPLICABLE.
Keep soft outcomes separate as APPLIED, ADAPTED, or SKIPPED; soft choices do not block submission readiness.
Separate
```
substance
```
problems from
```
misread-risk
```
problems. Do not recommend new experiments for wording problems that can be fixed by narrowing or clarifying a claim.
Diagnose concrete prose defects without using AI-detector, perplexity, authenticity, or human-likeness scores as compliance or readiness evidence; never infer AI authorship from style alone.
Compare every Anti-AI revision with its source for changes in claim strength, causality, scope, terminology, citations, values, and caveats.
When a characterized systematic error exists, verify the disclosed bias direction against the error mechanism and affected claims; require explicit indeterminacy when direction is unsupported.
When evidence sources have different roles, verify that validation, calibration, control, case-study, and exploratory sources never silently support stronger substantive claims.
Check whether the section and RQ order follows a coherent argumentative spine; treat this as adaptive structure, not a rejection criterion by itself.
For load-bearing threats, require an affected claim or RQ, a supported mitigation or explicit absence, and a residual boundary after mitigation.
Treat Related Work as a high-risk area. Enforce a comparison table only when
```
RELATED.COMPARISON_REQUIRED
```
is active; otherwise evaluate whether the comparison structure is adequate without imposing house style.
Enforce traditional concise headings only when
```
STRUCT.TRADITIONAL_HEADINGS
```
is active. In formal reviews of other work, criticize headings only when they cause a concrete clarity or venue problem.
For figures and tables, diagnose the issue and hand off to
```
paper-figures-tables
```
when creation, redesign, or LaTeX table work is needed.
For prose rewrites, diagnose the issue and hand off to
```
paper-writing
```
when drafting is the main task.
For rebuttals, answer every reviewer concern or explicitly mark it as deferred with a reason.

每一条批评意见都必须指向手稿中的具体位置、审稿人引用内容、本地制品或
```
MISSING
```
标记。
硬性结果仅使用PASS、FAIL、UNVERIFIED、NOT_APPLICABLE或WAIVED状态。绝不能将缺失的证据判定为PASS。
仅当手稿提供精确的制品、定位器和矛盾证据时，才记录Agent语义/手动FAIL状态。绝不能使用Agent记录来判定PASS、WAIVED或NOT_APPLICABLE。
将软性结果单独标记为APPLIED、ADAPTED或SKIPPED；软性选择不会阻碍投稿就绪。
将
```
substance
```
（实质性）问题与
```
misread-risk
```
（误读风险）问题区分开。对于可通过缩小或明确主张来解决的表述问题，请勿建议开展新实验。
诊断具体的文稿缺陷时，请勿将AI检测器、困惑度、真实性或类人度评分作为合规性或就绪性证据；绝不能仅从风格推断AI作者身份。
将每一项反AI修订内容与其原文对比，检查主张强度、因果关系、范围、术语、引文、数值和警告内容的变化。
当存在已确认的系统性错误时，验证披露的偏差方向是否与错误机制和受影响的主张相符；若方向无依据，需明确标注不确定性。
当证据来源具有不同作用时，验证验证、校准、控制、案例研究和探索性来源是否从未隐性支持更强的实质性主张。
检查章节和研究问题（RQ）的顺序是否遵循连贯的论证主线；将此视为适应性结构，而非单独的拒稿标准。
对于影响核心结论的威胁，要求明确受影响的主张或研究问题、有依据的缓解措施或明确说明未采取缓解措施的原因，以及缓解后的剩余边界。
将相关工作视为高风险领域。仅当
```
RELATED.COMPARISON_REQUIRED
```
激活时，才强制要求使用对比表格；否则评估对比结构是否充分，不强加内部格式要求。
仅当
```
STRUCT.TRADITIONAL_HEADINGS
```
激活时，才强制要求使用传统简洁标题。在正式评审他人工作时，仅当标题导致明确的清晰度问题或不符合投稿会议/期刊要求时，才对标题提出批评。
对于图表，诊断问题后，若需要创建、重新设计或LaTeX表格处理，请转至
```
paper-figures-tables
```
技能。
对于文稿重写，诊断问题后，若主要任务是撰写，请转至
```
paper-writing
```
技能。
对于反驳内容，需回应每一条审稿人关注的问题，或明确标记为延迟处理并说明原因。

Scripts

脚本

Use
```
scripts/count_rebuttal_limit.py
```
when a rebuttal, author response, or per-reviewer reply has a character or word limit.
For own-paper compliance audits, use the sibling
```
paper-policy/scripts/assess_compliance.py
```
; it reads artifacts and evidence without mutating them.
For project-local first passes, inspect the sibling runner's section-aware
```
soft-review-worklist.yaml
```
and locator-rich unused-key report; neither output is compliance evidence or mutation authorization.

当反驳内容、作者回复或针对单个审稿人的回复有字符或字数限制时，使用
```
scripts/count_rebuttal_limit.py
```
。
对于自有论文的合规性审核，使用关联技能
```
paper-policy/scripts/assess_compliance.py
```
；该脚本会读取制品和证据但不会对其进行修改。
对于项目本地的初步评审，检查关联运行器中基于章节的
```
soft-review-worklist.yaml
```
和包含定位信息的未使用关键词报告；这两个输出均不属于合规性证据或修改授权。

Reference Index

参考索引

```
references/review-modes.md
```
: mode selection and output shapes.
```
references/policy-compliance.md
```
: policy resolution, evidence statuses, waivers, and submission readiness.
```
references/pre-submission-audit.md
```
: manuscript audit dimensions and severity levels.
```
references/reviewer-panel.md
```
: reviewer personas, fairness firewall, and formal-review mode.
```
references/issue-board.md
```
: atomic issue schema for audits and external reviews.
```
references/rebuttal-strategy.md
```
: strategy selection, reviewer intent, and evidence planning.
```
references/rebuttal-drafting.md
```
: author-response structure, tone, and length discipline.
```
references/revision-plan.md
```
: mapping rebuttal promises to manuscript edits.
```
references/quality-gates.md
```
: final gates before returning a review or rebuttal artifact.

references/citation-and-evidence-policy.md

: local-only citation and evidence policy.

```
references/review-modes.md
```
：模式选择和输出形式。
```
references/policy-compliance.md
```
：政策解析、证据状态、豁免条款和投稿就绪情况。
```
references/pre-submission-audit.md
```
：手稿审核维度和严重程度等级。
```
references/reviewer-panel.md
```
：审稿人角色、公正性防火墙和正式评审模式。
```
references/issue-board.md
```
：审核和外部评审的原子化问题schema。
```
references/rebuttal-strategy.md
```
：策略选择、审稿人意图和证据规划。
```
references/rebuttal-drafting.md
```
：作者回复的结构、语气和篇幅准则。
```
references/revision-plan.md
```
：将反驳承诺映射到手稿修改内容。
```
references/quality-gates.md
```
：返回评审或反驳制品前的最终检查项。
```
references/citation-and-evidence-policy.md
```
：仅本地使用的引文和证据政策。