Skip to content
Back to search

paperjury

Pre-submission CS-conference LaTeX paper editing and adversarial review. Use for direct edits such as polish, de-AI, translating Chinese notes to LaTeX, compressing passages, captions, and experiment prose; for review or critique requests such as review, mock-review, 审稿, 评审; and for explicit auto or goal-driven review-revise loops. Runs deterministic Node guards, a durable ledger, isolated reviewer and juror fan-out, author sign-off, and compile or compliance checks. Not for writing a paper from scratch, figure generation, or official rebuttals.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/spark-to-paper-skills--paperjury-codex/badge.svg)](/listings/spark-to-paper-skills--paperjury-codex)

HTML

<a href="/listings/spark-to-paper-skills--paperjury-codex"><img src="/api/listings/spark-to-paper-skills--paperjury-codex/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

Plugin marketplace

codex marketplace add Spark-To-Paper-Skills/paperjury-codex

然后在 Codex plugin UI 里安装 **PaperJury Codex**。

稳定 pinned release:

codex marketplace add Spark-To-Paper-Skills/paperjury-codex@v1.0

每次 PaperJury workflow 开始时,插件会检查 GitHub 上的稳定 release tag。如果发现新版,只会给出 latest channel 和 pinned release 的安装命令,不会打断当前工作;如果 GitHub 不可达,就静默继续。设置 `PAPERJURY_DISABLE_UPDATE_CHECK=1` 可以关闭提醒。重新安装或切换 release channel 后,建议新开一个 Codex thread,让新版 skill 内容重新加载。

Legacy skill 安装

如果当前 Codex 界面还没有 plugin 安装入口,可以 clone 成 skill:

git clone https://github.com/Spark-To-Paper-Skills/paperjury-codex "$env:USERPROFILE\.codex\skills\paperjury"

**给 Codex / 编码 agent:** 先读 [`codex/AGENT-GUIDE.md`](codex/AGENT-GUIDE.md) 和 [`codex/runtime.md`](codex/runtime.md),再用 [`codex/phase-contracts.md`](codex/phase-contracts.md) 确定各阶段输入、输出、隔离和校验规则。Codex 版通过显式授权的 subagents 并行运行这套流程。

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 281 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. ---
  2. name: paperjury
  3. description: Pre-submission CS-conference LaTeX paper editing and adversarial review. Use for direct edits such as polish, de-AI, translating Chinese notes to LaTeX, compressing passages, captions, and experiment prose; for review or critique requests such as review, mock-review, 审稿, 评审; and for explicit auto or goal-driven review-revise loops. Runs deterministic Node guards, a durable ledger, isolated reviewer and juror fan-out, author sign-off, and compile or compliance checks. Not for writing a paper from scratch, figure generation, or official rebuttals.
  4. ---
  5. # PaperJury (CS-conference paper review and editing)
  6. PaperJury edits and hardens any CS-conference paper. It runs in
  7. three modes. In **direct-edit mode** (the common case) the user describes a change
  8. in Chinese or English and the LaTeX is edited directly through a CS-venue writing
  9. toolkit, with author sign-off. In **review mode** (occasional, pre-submission) it
  10. exposes the manuscript to a harsh, multi-perspective courtroom review engine that
  11. adjudicates each issue (N holistic domain reviewers -> contestability routing ->
  12. two-sided trial -> three-way verdict, with a polish track and a clerk-converged
  13. multi-round loop), gates every change behind consensus, and tracks issues in a durable
  14. ledger. In **auto mode** (unattended, opt-in via `/goal`) it runs that same engine
  15. toward a verifiable goal, applying safe fixes under a drift-bounded policy and
  16. queueing the risky ones for one human pass on return. All modes share the same
  17. writing toolkit, hard rules, ledger, and author sign-off (auto via up-front policy
  18. sign-off plus the queue, see hard rule 1).
  19. This skill is **fully generic**. It ships no hardcoded paths, no project files,
  20. and no embedded paper. Everything specific to a given paper (where the
  21. manuscript is, the venue, who signs off, the house style) is resolved at runtime
  22. or supplied by a config the *project* owns. The skill itself is the backbone;
  23. any concrete paper is just an instantiation of it.
  24. Scope: CS conferences only. Three venue families, each with its own style profile:
  25. - **Vision**: CVPR, ICCV, ECCV, WACV
  26. - **NLP**: ACL, EMNLP, NAACL, COLING
  27. - **ML**: ICLR, NeurIPS, ICML, AAAI, COLM
  28. ## Codex runtime
  29. This repository is the Codex-first PaperJury port. Read `codex/runtime.md`
  30. before running `review` or `auto`, then use `codex/phase-contracts.md` for phase
  31. inputs, outputs, isolation, and validation. Run deterministic guards with
  32. `node scripts/*.js`, and run semantic fan-out via Codex subagents only when the
  33. user has explicitly authorized multi-agent work.
  34. Soft update reminder: at the start of each PaperJury invocation, before choosing
  35. the mode or editing a manuscript, run `node scripts/check-update.js` from the
  36. skill root unless `PAPERJURY_DISABLE_UPDATE_CHECK=1` is set. If it reports an
  37. available update, show the notice once and continue. If the check is skipped,
  38. silent, or cannot reach GitHub, continue without mentioning it; update checks are
  39. never allowed to block review or editing.
  40. If host capabilities conflict with this file, keep the scientific invariants:
  41. author sign-off, reviewer isolation, no silent drops, a machine ledger, and honest
  42. degradation when a check cannot be verified.
  43. ## When to use / when not
  44. Three modes, one skill. Pick by what the user is asking for:
  45. - **Direct-edit mode (the common case).** The user describes a change in Chinese
  46. (or English) and wants the LaTeX edited directly: "把这段改成...", "polish this
  47. paragraph", "把我对 intro 的想法写成 LaTeX", "tighten this". No review panel; go
  48. straight to drafting the patch through the writing toolkit, with author sign-off.
  49. - **Review mode (occasional, pre-submission).** The user wants the paper critiqued
  50. or hardened: review / critique / 审稿 / 评审 / mock-review, or iterating a draft
  51. to clear reviewer-raised issues. This runs the courtroom review engine
  52. (`references/review-engine-v3.md`).
  53. - **Auto mode (unattended).** The user opts in via `/goal` (or config `mode: auto`)
  54. to run the review-revise loop AFK toward a verifiable goal. Establish the spine
  55. up front (the one human step), then the engine applies safe fixes under the
  56. bounded-aggressive policy and queues the rest. See `references/auto-mode.md`.
  57. Never self-detect auto; it is explicit only.
  58. Do NOT use for: writing a paper from scratch (use `ml-paper-writing`), figure or
  59. diagram generation (use `academic-plotting`), or an official-venue rebuttal (this
  60. is a pre-submission self-hardening loop, no score gate).
  61. ## The three primitives
  62. This paradigm is expressed as **Skill + Semantic Fan-Out + Memory**. Each carries
  63. one concern; together they replace the heavy per-round file-and-flag machinery a
  64. hand-rolled version accumulates.
  65. 1. **Skill (this folder) = entry point + methodology.** The protocol, the
  66. reviewer panel, the contestability routing, the writing toolkit, the human gates.
  67. Detail in `references/review-engine-v3.md`, `references/reviewer-personas.md`,
  68. `references/writing-toolkit.md`.
  69. 2. **Semantic fan-out = reviewer, jury, merge, audit, and drafting agents.** In
  70. Codex it is implemented by the orchestrator spawning isolated subagents, or by
  71. an explicitly labeled degraded single-agent path when subagents are unavailable.
  72. The v3 courtroom engine is
  73. `assign-reviewers` -> `reading-check` -> `coverage-auditor` -> `merge` ->
  74. {`trial` (+ escalate) || `polish`} -> `recall-audit` -> `drafter` ->
  75. {`edit-audit` | `meaning-audit`} -> `clerk`. Protocol + every orchestrator seam:
  76. `references/review-engine-v3.md`; Codex runtime mapping:
  77. `codex/runtime.md`; Codex-native phase contracts:
  78. `codex/phase-contracts.md`.
  79. 3. **Memory = durable state + learned conventions.** Two layers:
  80. - **Ledger** (`LEDGER.json` resolved at runtime = the machine source of truth,
  81. plus a rendered `LEDGER.md` view; managed by `scripts/ledger.js`): the live,
  82. mutable issue state across rounds and sessions. Schema + status state machine:
  83. `references/ledger-schema.md`.
  84. - **Project conventions**: stable house-style and venue conventions stored in
  85. the active project, for example `.paper-review/CONVENTIONS.md` or the host's
  86. native project memory. Do not store transient issue state outside the ledger.
  87. ## Resolving inputs at runtime (no hardcoded paths)
  88. The skill ships ZERO hardcoded paths or project files. On trigger it resolves
  89. each input by **discovery first, then asking**:
  90. - **manuscript**: detect the main source (the `.tex` with `\documentclass` /
  91. `\begin{document}`, or the file the user names). If several candidates, ask.
  92. - **venue_family**: the user can name it, or an agent reads the class file to
  93. GUESS the family (e.g. a cvpr/iccv style, an acl style, a neurips/iclr style).
  94. There is no hardcoded venue list and no deterministic detector; if unclear, ask.
  95. - **ledger**: default to `<manuscript-dir>/.paper-review/LEDGER.json` (the machine
  96. source of truth; `scripts/ledger.js` also renders a `LEDGER.md` view). Create if
  97. absent, reuse if present. The user may point elsewhere.
  98. - **author**: ask who signs off on edits (default: the current user). Every edit
  99. needs explicit authorization.
  100. - **personas**: default to N domain-expert holistic reviewers assigned at runtime
  101. (`assign-reviewers`, from the project gatekeeper core + a generated domain overlay);
  102. the three generic lenses in `references/reviewer-personas.md` are the degrade
  103. fallback. If the project defines its own named reviewer subagents, use them as
  104. `agentType`; otherwise inline the persona prompts.
  105. - **style_profile**: start from the venue-family default; refine from any
  106. conventions recalled from memory or pinned in a project config.
  107. A project MAY pin these by dropping a config in ITS OWN repo (see
  108. `configs/config-template.md` for the shape). That file is owned by the project,
  109. never by this skill. At round start, recall any pinned conventions from memory.
  110. ## Direct-edit mode (the common case)
  111. The user states a change in Chinese or English; you draft and apply the LaTeX edit.
  112. No panel, no ledger, no discussion. Minimal flow:
  113. 1. **Locate.** Resolve the manuscript and find the target passage the instruction
  114. refers to (a paragraph, sentence, caption, table cell). If it is ambiguous on a
  115. large file, ask which passage; do not guess.
  116. 2. **Draft.** Pick the writing-toolkit prompt matching the instruction
  117. (`translate-to-english` for a Chinese idea, `polish-english` / `de-ai` for a
  118. rewrite, `compress` / `expand` for length, `caption` / `experiment-analysis`
  119. for those units) and draft the LaTeX patch to do exactly what was asked. The
  120. Common guards apply (LaTeX-safe, plain CS prose, no log leakage into the .tex).
  121. 3. **Self-gate.** Run `logic-check` on the drafted passage.
  122. 4. **Sign-off.** Show the patch and get explicit author approval (hard rule 1).
  123. 5. **Apply.** Write only the patch into the manuscript; keep any back-translation
  124. or note author-side.
  125. This is the writing toolkit used on its own. Escalate to review mode only when the
  126. user wants the paper critiqued or hardened, not for a single asked-for edit.
  127. ## Why fan-out is separate from conversation
  128. The reviewer panel and the trial jury are pure fan-out: spawn, collect,
  129. merge. Codex uses subagents when the user has explicitly authorized parallel
  130. agent work. Each reviewer or juror gets only the quoted manuscript context and a
  131. strict JSON contract; no peer report, ledger, prior round, or project file path is
  132. included in the prompt.
  133. But the loop has genuine human gates (the author reviews the issue list, gives
  134. per-issue direction, authorizes edits, breaks ties). Fan-out phases run to
  135. completion and return a result; they do not pause mid-run for hours of human
  136. input. So:
  137. - fan-out steps (reviewers, trial, polish, recall, merge) -> **Codex subagents**
  138. - human gates (per-issue direction, authorization, override) -> **main conversation turns**
  139. - cross-round truth (the ledger) + stable conventions -> **project-owned memory files or host memory**
  140. ## Review mode: one round, end to end
  141. The full adversarial loop (the v3 courtroom engine). Use it to harden the paper,
  142. not for a single asked-for edit (that is direct-edit mode). Full protocol + the
  143. 14 orchestrator seams: `references/review-engine-v3.md`. `[SF]` = semantic
  144. fan-out step using Codex subagents per `codex/runtime.md` and
  145. `codex/phase-contracts.md`; `[det]` = deterministic Node guard run
  146. orchestrator-side between semantic steps; `[HUMAN]` = author gate; `[LEDGER]` =
  147. state write.
  148. 1. **Resolve + recall.** Resolve the inputs above; recall this paper's conventions
  149. from memory. Pick scope: `full` (whole paper) or `passage` (one section / para / claim).
  150. 2. **`[det]` decompose.** Split the manuscript into reading units + stable
  151. `passage_id`s + the canonical section list.
  152. 3. **`[SF]` assign-reviewers** + **`[HUMAN]` confirm.** Name N subfields (2-4,
  153. default 3); instantiate N holistic domain reviewers from the gatekeeper core + a
  154. generated overlay. An unconfirmable slot degrades per slot to a generic gatekeeper
  155. (the three generic lenses in `reviewer-personas.md` are the fallback). The author
  156. confirms the assignment (or pins it via config).
  157. 4. **`[SF]` reading-check.** Each reviewer reads the WHOLE paper → weaknesses
  158. {`significance`(major|minor), `kind`(mechanical|substantive), verbatim quote —
  159. cannot quote = did not read} + one `overall_confidence` + a per-section coverage
  160. report. Anti-skim is three layers: `[det]` per-section quote-verify, `[SF]`
  161. coverage-auditor, `[SF]` targeted re-invoke.
  162. 5. **`[SF]` merge.** Semantic dedup across reviewers; derive `significance` (MAX) /
  163. `kind` (substantive-dominates) / corroboration. `[LEDGER]` intake as `raised`.
  164. 6. **`[det]` route.** mechanical → polish; substantive&minor → polish;
  165. substantive&major → trial (two parallel tracks).
  166. 7. **`[SF]` trial.** Per substantive-major charge: a whole-paper DEFENSE → 5
  167. decorrelated local-context jurors (+ on-demand expansion) → a deterministic verdict
  168. (decide iff quorum `surviving >= ceil(0.8*jurySize)` AND one side `> 60%` of
  169. surviving votes; else escalate to 12). Verdict ∈ {invalid-drop, valid-fixable,
  170. author-required, escalate}; the judge sets a `close_criterion` ONLY for a
  171. valid-fixable charge, satisfiable by editing existing text (no new data). `[SF]`
  172. polish runs the off-gate mechanical/minor track in parallel (never silently dropped).
  173. 8. **`[SF]` recall-audit.** Mode A revives wrongly-dropped charges; Mode B spot-checks
  174. strong-consensus majors BEFORE the edit. Runs before the drafter.
  175. 9. **`[HUMAN]` Authorize + `[SF]` drafter + edit-safety.** On authorization, the
  176. drafter writes the minimal patch per surviving valid-fixable. The edit-safety chain
  177. gates it: `[det]` anchor-diff + cross-ref → `[SF]` meaning-audit (frozen anchor,
  178. four-state) / edit-audit (risky non-anchor); `[det]` apply-patch + compile-guard land
  179. a passing patch and `[LEDGER]` mark `closed`; a drift / anchor / failed edit is
  180. reverted and queued. Revision logs / back-translations stay author-side.
  181. 10. **`[SF]` clerk + report.** The clerk reconciles the round boundary (carried
  182. open-questions vs this round's edits, via a passage_id + similarity merge key) and
  183. emits convergence counts. Summarize new/closed counts; in review mode do not
  184. auto-start the next round (auto mode drives the outer loop via `/goal`).
  185. GATE: `node scripts/ledger.js gate` = 0 gate-blocking active major (gate-blocking =
  186. {raised, in-trial, re-trial, valid-fixable}; author-required / queued / dropped /
  187. closed are gate-OK and author-required accumulates to the queue). Full protocol +
  188. ledger schema + status machine: `references/review-engine-v3.md`,
  189. `references/ledger-schema.md`. The single-pass 3-reviewer panel is available as
  190. the `review-panel` fast path in `codex/phase-contracts.md`.
  191. ## Hard rules (load-bearing, venue-agnostic)
  192. 1. **Never edit the manuscript without explicit author sign-off.** Auto-mode
  193. carve-out: the rule HOLDS; auto satisfies it via UP-FRONT sign-off (the spine
  194. confirmation + the pre-authorized bounded-aggressive policy) plus the return
  195. queue, not per-edit sign-off. Nothing outside the authorized envelope is applied.
  196. 2. **Reviewers / jurors are isolated.** Fresh eyes per round: no cross-talk, no
  197. prior-round leakage, no sight of the ledger. Enforced by (a) what goes into each
  198. agent's prompt AND (b) an explicit ISOLATION instruction in every reviewer-type
  199. prompt telling the agent to judge only the quoted text and not read files.
  200. 3. **A valid-fixable issue carries a `close_criterion`** (one concrete sentence an
  201. edit must satisfy), set by the judge at trial; it is null at intake.
  202. 4. **No leakage into the reviewed text.** Revision logs, back-translations, and
  203. self-check verdicts are author-side aids; they never enter the manuscript or
  204. any frozen snapshot.
  205. 5. **Disagreement resolves through discussion, then override** (logged), never a
  206. silent dismissal.
  207. 6. **No hardcoded paths or project files in the skill.** Resolve at runtime.
  208. ## Memory convention
  209. - At round start: recall the paper's conventions (house style, venue, persona
  210. tuning) from memory; read the resolved `LEDGER.json` for open issues.
  211. - During the round: the ledger is the only mutable truth; update it at merge,
  212. trial verdicts, recall, and close.
  213. - After the round: persist any newly learned stable convention to memory (e.g. a
  214. house-style rule a reviewer surfaced), not the transient issue state.
  215. ## Intensity and host-specific launch notes
  216. The quick panel can run in stronger or cheaper forms using the `review-panel`
  217. fast path in `codex/phase-contracts.md`:
  218. - **loop-until-dry**: re-runs independent fresh panels and accumulates only issues
  219. not seen before, stopping after `dryStop` consecutive passes that add no
  220. surviving issue (hard cap `maxRounds`). Raises recall past a single pass.
  221. - **adversarial verify**: each new issue faces perspective-diverse skeptics
  222. (misreading / already-addressed / scope-or-severity) and is kept unless a
  223. majority refute it, filtering plausible-but-wrong issues before they reach the
  224. ledger. Bias is to keep, so real flaws are not lost.
  225. Toggle via args: thorough mode -> defaults (`maxRounds` 4, `dryStop` 2,
  226. `verify` true); light mode -> pass `{maxRounds:1, verify:false}` for the basic
  227. single-panel form. The loop is budget-aware and stops early if the token budget
  228. runs low. Codex treats intensity as an ordinary runtime setting, not as a
  229. separate launch keyword.
  230. ## Capabilities and status
  231. Built: the review engine; the submission-readiness checker (deterministic desk-reject screening plus a real LaTeX compile, degrading to a structural lint when no toolchain is present); and auto mode (the review-revise loop toward a goal under a drift-bounded policy, applying safe fixes and queueing risky ones for author review). Roadmap: vision-based layout verification, automatic venue detection from the class file, and reviewer personas tuned to each venue community.
  232. ## Related skills
  233. - `ml-paper-writing`: from-scratch drafting, citation verification (never
  234. hallucinate citations), conference checklists. This loop borrows its
  235. sentence-level guidance for the edit-drafting step rather than duplicating it.
  236. - `academic-plotting`: figure and architecture-diagram generation (out of scope
  237. here; this loop edits text and captions, not figure images).
README.md · README · 296 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. **中文** · [English](README.en.md)
  2. <p align="center">
  3. <img src="docs/paperjury-mark.png" alt="PaperJury logo" width="120">
  4. </p>
  5. <h1 align="center">PaperJury Codex</h1>
  6. <p align="center">真正投稿前,先让 AI reviewer 把该挑的坑挑出来。</p>
  7. <p align="center">
  8. <a href="https://arxiv.org/abs/2606.16322"><img alt="阅读论文(arXiv)" src="https://img.shields.io/badge/arXiv-2606.16322-b31b1b?style=for-the-badge&logo=arxiv&logoColor=white"></a>
  9. <a href="https://spark-to-paper-skills.github.io/paperjury/overview.html?lang=zh"><img alt="打开在线交互式总览" src="https://img.shields.io/badge/在线交互式总览-d6a14b?style=for-the-badge&logo=githubpages&logoColor=white"></a>
  10. <a href="https://github.com/Spark-To-Paper-Skills/paperjury"><img alt="Claude Code 版本" src="https://img.shields.io/badge/Claude_Code_version-2b2d42?style=for-the-badge"></a>
  11. <a href="samples/dogfood/"><img alt="查看真实样例" src="https://img.shields.io/badge/真实样例-Dogfood-2f7d55?style=for-the-badge"></a>
  12. <a href="https://github.com/Spark-To-Paper-Skills/paperjury-codex/stargazers"><img alt="Star this repo" src="https://img.shields.io/badge/GitHub-Star-3b3d47?style=for-the-badge&logo=github&logoColor=white"></a>
  13. <a href="https://github.com/Spark-To-Paper-Skills/paperjury-codex/releases"><img alt="Open releases" src="https://img.shields.io/badge/Releases-Open-3b3d47?style=for-the-badge"></a>
  14. <img alt="License: MIT" src="https://img.shields.io/badge/license-MIT-3b3d47?style=for-the-badge">
  15. </p>
  16. <p align="center">
  17. <a href="https://spark-to-paper-skills.github.io/paperjury/overview.html?lang=zh">
  18. <picture>
  19. <source media="(prefers-color-scheme: dark)" srcset="docs/overview-card-dark.png">
  20. <source media="(prefers-color-scheme: light)" srcset="docs/overview-card-light.png">
  21. <img src="docs/overview-card.png" alt="PaperJury 交互式总览" width="760">
  22. </picture>
  23. </a>
  24. </p>
  25. > **一句话版:** PaperJury Codex 会在投稿前替你的论文做一轮“刻薄但有证据”的 AI 预审:先找 reviewer 可能会抓住的问题,再判断这些问题到底成不成立;能安全修改的只给最小补丁,缺实验、缺证据的交回作者,不靠谱的意见直接驳回。
  26. 刚写完初稿、投稿前想再自查一遍,或者只想先让它挑一遍风险,都可以直接用日常话描述需求;不用先研究它的内部机制。
  27. ## News
  28. - 🎉 **RedNote(小红书)里程碑:** 相关分享已经达到 **3 万浏览**、**1.8k 收藏**。感谢大家转发和收藏,也感谢大家把 PaperJury 推荐给更多正在赶论文、改论文的朋友。
  29. - 📄 **PaperJury 论文已上 arXiv。** arXiv 页面:[*PaperJury: Due-Process Review for Bounded LaTeX Revision*](https://arxiv.org/abs/2606.16322)(arXiv:2606.16322)。论文系统介绍了「审稿 → 裁定 → 修改 → 复查」这套引擎。
  30. - 🚀 **Codex plugin 已发布。** PaperJury Codex 现在可以通过 Codex plugin marketplace 安装;需要时也可以继续用 legacy clone 安装。
  31. - 🔔 **v1.0 release。** Codex plugin 带有软更新提醒;发现新的稳定 tag 时只提示,不打断当前工作。
  32. - 🧪 **Dogfood sample 已加入。** 仓库里放了修改前后 PDF 和人工核对过的运行报告。
  33. ---
  34. **PaperJury Codex 是 PaperJury 的 Codex 版。** 它把投稿前自查组织成一套闭环:**审稿 → 裁定 → 修改 → 复查**。你可以让它像 reviewer 一样挑问题,也可以让它安全地改一处 LaTeX / Markdown;真正落稿前,它会先把补丁交给你确认。
  35. 它不会照单全收 AI 反馈,而是先把每条意见分成三类:
  36. | 结果 | 含义 |
  37. |---|---|
  38. | **✅ 安全修复** | 表达不清、claim 过强、结构不顺这类文本问题;不需要补实验,也不会把原意改偏。 |
  39. | **🧑‍💻 作者处理** | 缺实验、缺 ablation、缺数据或证据,必须由作者自己判断。 |
  40. | **🛑 不成立** | AI reviewer 误读了论文,或者提了不该改的问题。 |
  41. > [!IMPORTANT]
  42. > PaperJury 是投稿前的自查工具,**不能替代作者的科学判断,也不能替代 peer review**。它不能用来编造实验、伪造结果、添加没有证据支撑的 claim,或掩盖论文局限。遇到需要新实验、缺证据、需要作者私有知识或研究判断的问题,它都会交回作者处理。
  43. ## 适合谁
  44. | 你现在的情况 | 可以直接这样用 |
  45. |---|---|
  46. | **📝 刚写完初稿** | 让它像 reviewer 一样通读全文,先找最可能影响投稿的问题。 |
  47. | **🔍 投稿前最后自查** | 让它检查 claim 是否说过头、实验是否撑得住,以及有没有明显格式风险。 |
  48. | **✍️ 只想改一段话** | 直接说「把这段改紧一点,但不要改变 claim」,它会先起草补丁,等你确认后再改。 |
  49. | **🔁 想多轮打磨但不想一直盯着** | 明确授权 auto 模式;安全修改可以落稿,高风险问题仍会交回作者决定。 |
  50. ## 你会得到什么
  51. | 输出 | 内容 |
  52. |---|---|
  53. | **📋 问题清单** | 每条 reviewer-style 问题都会带证据、位置、判断结果和当前状态;不会把一堆意见直接倒进正文。 |
  54. | **🧩 可审阅补丁** | 只有安全修复会进入最小补丁;高风险改动会先放着,等作者决定。 |
  55. | **🛠️ 复查报告** | 有 LaTeX 工具链就真实编译;没有就明说哪些检查做不了,不会假装验证过。 |
  56. | **🧪 真实样例** | [`samples/dogfood/`](samples/dogfood/) 里有修改前后 PDF 和人工核对过的运行报告。 |
  57. ---
  58. ## 目录
  59. - [News](#news)
  60. - [适合谁](#适合谁)
  61. - [你会得到什么](#你会得到什么)
  62. - [快速上手](#快速上手)
  63. - [能帮你做什么](#能帮你做什么)
  64. - [三种模式](#三种模式)
  65. - [安装](#安装)
  66. - [Codex 版和 Claude Code 版](#codex-版和-claude-code-版)
  67. - [真实跑一遍](#真实跑一遍)
  68. - [深入了解](#深入了解)
  69. - [Roadmap](#roadmap)
  70. - [致谢](#致谢)
  71. ---
  72. ## 快速上手
  73. 先把 marketplace 加到 Codex:
  74. ```powershell
  75. # Latest channel:安装或重新安装时使用当前 main 分支
  76. codex marketplace add Spark-To-Paper-Skills/paperjury-codex
  77. # 稳定 pinned release:可复现的 v1.0 安装
  78. codex marketplace add Spark-To-Paper-Skills/paperjury-codex@v1.0
  79. ```
  80. 然后在 Codex plugin UI 里安装 **PaperJury Codex**。安装后可以直接在论文项目里说:
  81. ```text
  82. 审稿,重点看实验设计、claim 强度和格式风险。
  83. ```
  84. 或者:
  85. ```text
  86. 把这段 intro 改紧一些,但不要改变原来的 claim。
  87. ```
  88. 安装后会出现 `paperjury` skill,随插件一起安装的还有确定性 Node guards。`node` 是必需的;LaTeX 工具链可选,有工具链时会真实编译,没有时会明说哪些检查做不了。
  89. ## 能帮你做什么
  90. | 场景 | PaperJury Codex 会怎么做 |
  91. |---|---|
  92. | **🔎 投稿前挑问题** | 模拟几位不同方向的 reviewer 通读全文,找出真正可能被抓住的弱点,并把致命问题和小修小补分开。 |
  93. | **✍️ 安全改 LaTeX / Markdown** | 只针对你指定的位置起草补丁,自检后再交给你确认;不会把一处小改扩成整篇重写。 |
  94. | **🛡️ 复查格式风险** | 本机有 LaTeX 工具链时会真实编译,检查报错、未定义引用、overfull box、页数和常见 desk-reject 风险;没有工具链时会明说。 |
  95. | **🔁 多轮打磨** | 在你明确授权的 auto 模式下,多轮跑完「评审-修订-复查」;安全修改可以自动应用,高风险问题会留给作者处理。 |
  96. ## 三种模式
  97. | 模式 | 什么时候用 | 行为 | 人工确认 |
  98. |---|---|---|---|
  99. | **✍️ direct-edit**(常用) | 只想改一处文字、caption、LaTeX 表达或段落结构。 | 不开评审面板,直接用写作工具包起草补丁。 | 作者确认后再应用。 |
  100. | **🔎 review**(偶尔) | 想让它审稿、挑问题、mock-review,或只审某一节 / 某条 claim。 | 启动对抗式评审引擎,先判断问题是否成立,再决定要不要修改。 | 每处改动逐一确认。 |
  101. | **🔁 auto**(无人值守) | 已经明确给出 Codex goal 或配置 `mode: auto`,希望它多轮跑到一个可验证目标。 | 先确认 `spine` 和评审分配,再按 bounded-aggressive + edit-safety 策略迭代。 | 先给整体授权;高风险项仍交回作者。 |
  102. 简单说:**改一处 → 直接说;想被挑刺 → 说「审稿」;想让它多轮跑 → 给它一个 Codex goal。**
  103. > [!WARNING]
  104. > **auto 必须明确开启。** 只打开工具权限再发普通 prompt,只会跑一轮就停,不会进入多轮循环。原因见 [`codex/AGENT-GUIDE.md`](codex/AGENT-GUIDE.md)。
  105. ## 安装
  106. ### Plugin marketplace
  107. ```powershell
  108. codex marketplace add Spark-To-Paper-Skills/paperjury-codex
  109. ```
  110. 然后在 Codex plugin UI 里安装 **PaperJury Codex**。
  111. 稳定 pinned release:
  112. ```powershell
  113. codex marketplace add Spark-To-Paper-Skills/paperjury-codex@v1.0
  114. ```
  115. 每次 PaperJury workflow 开始时,插件会检查 GitHub 上的稳定 release tag。如果发现新版,只会给出 latest channel 和 pinned release 的安装命令,不会打断当前工作;如果 GitHub 不可达,就静默继续。设置 `PAPERJURY_DISABLE_UPDATE_CHECK=1` 可以关闭提醒。重新安装或切换 release channel 后,建议新开一个 Codex thread,让新版 skill 内容重新加载。
  116. ### Legacy skill 安装
  117. 如果当前 Codex 界面还没有 plugin 安装入口,可以 clone 成 skill:
  118. ```powershell
  119. git clone https://github.com/Spark-To-Paper-Skills/paperjury-codex "$env:USERPROFILE\.codex\skills\paperjury"
  120. ```
  121. **给 Codex / 编码 agent:** 先读 [`codex/AGENT-GUIDE.md`](codex/AGENT-GUIDE.md) 和 [`codex/runtime.md`](codex/runtime.md),再用 [`codex/phase-contracts.md`](codex/phase-contracts.md) 确定各阶段输入、输出、隔离和校验规则。Codex 版通过显式授权的 subagents 并行运行这套流程。
  122. ## Codex 版和 Claude Code 版
  123. | 版本 | 入口 | 差异 |
  124. |---|---|---|
  125. | **Codex 版** | 本仓库;Codex plugin 或 `.codex/skills/` | 语义阶段由显式授权的 Codex subagents 运行;阶段契约写在 `codex/phase-contracts.md`;确定性检查由 orchestrator 通过 Node 调用。 |
  126. | **Claude Code 版** | [paperjury](https://github.com/Spark-To-Paper-Skills/paperjury) | 使用 Claude Code skill / plugin 路线;语义阶段由 Claude Workflow 文件驱动。 |
  127. 两个版本共享同一套方法论:bounded review、issue ledger、recall audit、edit-safety、compile guard 和 submission-readiness。Codex 版不携带也不执行 Claude Workflow 文件。
  128. ## 真实跑一遍
  129. 仓库里有一个 dogfood sample:在一篇真实草稿上跑完整多轮评审,附**修改前后 PDF** 和一份**人工核对过的运行报告**。
  130. [`samples/dogfood/`](samples/dogfood/)([`original_draft.pdf`](samples/dogfood/original_draft.pdf) · [`revised_draft.pdf`](samples/dogfood/revised_draft.pdf) · [运行报告](samples/dogfood/RUN_REPORT.zh-CN.md))
  131. 如果只想确认稿件不会先被格式问题挡住,可以说:
  132. ```text
  133. 跑一下 submission-readiness / 合规检查。
  134. ```
  135. 它会做确定性格式筛查,再配合编译驱动的版面检查。
  136. ## 深入了解
  137. 新用户可以先跳过这一节。想看机制、源码结构或 agent 驱动方式,可以从这里开始:
  138. | 你想了解 | 入口 |
  139. |---|---|
  140. | 真实运行效果 | [`samples/dogfood/RUN_REPORT.zh-CN.md`](samples/dogfood/RUN_REPORT.zh-CN.md) |
  141. | Codex runtime / agent 驱动方式 | [`codex/AGENT-GUIDE.md`](codex/AGENT-GUIDE.md) · [`codex/runtime.md`](codex/runtime.md) |
  142. | 语义阶段输入输出和隔离契约 | [`codex/phase-contracts.md`](codex/phase-contracts.md) |
  143. | 完整协议和状态机 | [`references/review-engine-v3.md`](references/review-engine-v3.md) · [`references/ledger-schema.md`](references/ledger-schema.md) |
  144. | 在线可视化说明 | [交互式总览](https://spark-to-paper-skills.github.io/paperjury/overview.html?lang=zh) |
  145. <details>
  146. <summary><b>展开机制、架构和项目结构说明</b></summary>
  147. ### 引擎总览
  148. PaperJury 把审稿拆成一套有边界的“庭审”流程:先由有限数量的 reviewer 找问题,再把有争议的意见拿出来审议;编辑阶段按风险加护栏,多轮结束时由确定性脚本判断是否收敛。
  149. ```text
  150. assign-reviewers → reading-check → coverage-auditor → merge
  151. → { trial ‖ polish } → recall-audit → drafter
  152. → { edit-audit | meaning-audit } → clerk
  153. ```
  154. ### 确定性步骤
  155. 1. **读稿分解**:把手稿切成阅读单元、规范段落列表和稳定段落编号,防止问题锚点漂移。
  156. 2. **核心声明**(仅 auto 模式):提取核心 claim,交给作者确认后冻结为配置。
  157. 3. **Ledger**:用机器可读的记录保存活跃问题,跨轮次、跨会话都能接上。只要没有仍然阻塞的 major 问题,就视为工具侧完成;author-required 会进入人工待办,不算工具侧未完成。
  158. 4. **日志**:编辑历史只追加记录,方便回滚。
  159. 5. **补丁应用**:原子性应用编辑,记录日志,必要时可以恢复。
  160. 6. **锚点追踪**:定位已冻结的核心 claim;上下文变化时,标出需要重新审计的部分。
  161. 7. **交叉引用检查**:编辑前先查改动关键词是否还出现在其他位置;如果出现,就标记为需要语义审计。
  162. 8. **编译检查**:尝试真实 LaTeX 编译;无法编译时退到结构检查,并明确说明哪些结果不可验证。
  163. 9. **提交合规检查**:用脚本先筛一遍常见提交格式风险。
  164. ### 语义步骤
  165. 1. **评审员分配**:根据论文研究方向,分配 N 位领域 reviewer。
  166. 2. **完整阅读检查**:每位 holistic reviewer 通读全文一遍,列出弱点、原文引文、总体置信度和按节覆盖情况;引不出原文,就视为没有真读。
  167. 3. **覆盖审计**:检查哪些 reviewer / section 组合可能被略读。
  168. 4. **去重**:合并重复评论,并整理重要性、问题类型和交叉确认情况。
  169. 5. **审议(trial)**:对有争议的问题开庭。先由 5 人审议,必要时升到 12 人;法官把成立的问题判定为 `valid-fixable` 或 `author-required`。
  170. 6. **润色**:快速路径处理机械性问题和轻微问题;如果判断不稳,就升级回审议。
  171. 7. **召回审计(recall)**:找回误判丢掉的问题,并在落稿前抽查强共识 major,防止集体误判。
  172. 8. **编辑起草**:对确认可修的问题起草最小改动。
  173. 9. **编辑审计 / 含义审计**:检查高风险改动、跨节一致性、冻结锚点和论证链条。
  174. 10. **书记官**:汇总本轮结果,合并重复项,整理残留问题,并用确定性规则判断是否收敛。
  175. 也支持简化的 3 人评审小组,作为快速路径。
  176. ## 架构说明
  177. - Codex 不执行 Claude Workflow 文件。语义阶段定义在 `codex/phase-contracts.md`;确定性检查由 orchestrator 通过 Node 调用。
  178. - `compile-guard.js` 不会假装验证过:无法真正编译时,就退到结构 lint,并报告 `compiled:null`。
  179. - 提交就绪检查分两部分:A = `compliance-check.js` + 一个语义 agent;B = 复用 `compile-guard.js` 的编译检查,再让模型读取 PDF 做版面复查。
  180. - 你的项目文件、ledger、journal 和 patch 都留在本地论文项目里。PaperJury 自己没有后端或服务器,也就不会把文件上传到 PaperJury 服务器。审稿走的是你自己的 Codex session;模型本身仍可能跑在云端,内容如何被处理取决于宿主环境的条款和设置,PaperJury 不会再额外加一层。
  181. ## 项目结构
  182. | 路径 | 作用 |
  183. |---|---|
  184. | `plugins/paperjury-codex/` | Codex plugin 发布包;README、skill、运行资源和 marketplace metadata 都在这里同步。 |
  185. | `codex/` | Codex runtime 映射、phase contracts 和 agent 驱动说明。 |
  186. | `agents/` | Codex 侧语义 agent 定义。 |
  187. | `scripts/` | 确定性检查脚本:ledger、journal、apply-patch、anchor-diff、cross-ref、compile-guard、doctor 等。 |
  188. | `references/` | 引擎协议、ledger schema、评审者人格、写作工具和方法论。 |
  189. | `docs/` | 交互式总览、站点资产和设计说明入口。 |
  190. | `samples/dogfood/` | 真实草稿的 before/after PDF 和人工核对过的运行报告。 |
  191. | `tests/` | 确定性脚本和核心状态机测试。 |
  192. </details>
  193. ## Roadmap
  194. - [x] **Codex plugin marketplace release。** 将 PaperJury 打包为可通过 Codex plugin marketplace 路线直接安装的版本,同时保留 legacy clone 安装。
  195. - [x] **软更新提醒。** 启动时检查有没有更新的稳定版 tag,有就给一条非阻塞提示。
  196. - [ ] **快速版本 / quick mode。** 更快、更省 token 的检查路径;不追求完整庭审深度,先给一轮可用的快速 triage。
  197. - [ ] **按不同会议的审稿口味调整 reviewer persona。** CVPR、ACL、NeurIPS 的 reviewer 关注点并不一样;目标是让评审更贴近各自社区的预期。
  198. - [ ] **基于视觉的版面校验。** 编译、渲染、再检查版面,不只看编译日志。
  199. - [ ] **从 `.cls` / 模板自动识别 venue。**
  200. - [ ] **用更多真实论文做规模化验证。**
  201. <details>
  202. <summary><b>更多文件与路径</b></summary>
  203. - 引擎协议:`references/review-engine-v3.md`
  204. - Codex 运行时映射:`codex/runtime.md`
  205. - Codex 语义阶段契约:`codex/phase-contracts.md`
  206. - 自动模式:`references/auto-mode.md`
  207. - 评审者角色、编辑工具:`references/reviewer-personas.md`、`references/writing-toolkit.md`
  208. - 账本结构和状态:`references/ledger-schema.md`
  209. - 提交合规:`references/submission-compliance.md`
  210. - Codex runtime / agent 说明:`codex/AGENT-GUIDE.md`
  211. - 脚本:`scripts/`
  212. - Codex 运行包:`codex/`
  213. </details>
  214. ## 致谢
  215. PaperJury 的 spine 和防漂移设计受 [PaperSpine](https://github.com/WUBING2023/PaperSpine) 启发,尤其是 anchor logic-transfer audit、claim register、minimal-edit 且保义的改写策略。PaperSpine 更偏 motivation-driven 的论文起草和改写;PaperJury 借用了其中的 anchoring 思路,以及“可检查步骤交给确定性脚本、判断交给 model agent”的分工,再在此基础上加入对抗式 review 和庭审式裁定流程。

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.