name: paperjury description: Three modes for CS-conference papers (CVPR/ICCV/ECCV vision, ACL/EMNLP/NAACL NLP, ICLR/NeurIPS/ICML/AAAI ML). DIRECT-EDIT mode (common): the user describes a change in Chinese or English and the manuscript (LaTeX or Markdown) is edited directly through a CS-venue writing toolkit with author sign-off (use for 改这段 / 把中文想法写成 latex / polish / de-AI / translate / compress a passage). REVIEW mode (occasional, pre-submission): harden the paper through an adversarial courtroom review engine (N holistic domain reviewers / contestability routing / two-sided trial / three-way verdict / clerk-converged multi-round loop) with consensus-gated, author-signed revisions (use for review / critique / 审稿 / 评审 / mock-review). AUTO mode (unattended, opt-in via /goal): run the review-revise loop toward a verifiable goal, applying safe fixes under a drift-bounded policy and queueing risky ones. Resolves all inputs at runtime, no hardcoded paths. Not a from-scratch drafter (use ml-paper-writing) and not an official-venue rebuttal. version: 1.2.1 author: Yiran Wang license: MIT tags: [Academic Writing, Peer Review, Adversarial Review, CVPR, ICCV, ECCV, ACL, EMNLP, NAACL, ICLR, NeurIPS, ICML, AAAI, Workflow, LaTeX]

PaperJury (CS-conference paper review and editing)

PaperJury edits and hardens any CS-conference paper. It runs in three modes. In direct-edit mode (the common case) the user describes a change in Chinese or English and the LaTeX is edited directly through a CS-venue writing toolkit, with author sign-off. In review mode (occasional, pre-submission) it exposes the manuscript to a harsh, multi-perspective courtroom review engine that adjudicates each issue (N holistic domain reviewers -> contestability routing -> two-sided trial -> three-way verdict, with a polish track and a clerk-converged multi-round loop), gates every change behind consensus, and tracks issues in a durable ledger. In auto mode (unattended, opt-in via /goal) it runs that same engine toward a verifiable goal, applying safe fixes under a drift-bounded policy and queueing the risky ones for one human pass on return. All modes share the same writing toolkit, hard rules, ledger, and author sign-off (auto via up-front policy sign-off plus the queue, see hard rule 1).

This skill is fully generic. It ships no hardcoded paths, no project files, and no embedded paper. Everything specific to a given paper (where the manuscript is, the venue, who signs off, the house style) is resolved at runtime or supplied by a config the project owns. The skill itself is the backbone; any concrete paper is just an instantiation of it.

Scope: CS conferences only. Three venue families, each with its own style profile:

• Vision: CVPR, ICCV, ECCV, WACV
• NLP: ACL, EMNLP, NAACL, COLING
• ML: ICLR, NeurIPS, ICML, AAAI, COLM

When to use / when not

Three modes, one skill. Pick by what the user is asking for:

• Direct-edit mode (the common case). The user describes a change in Chinese (or English) and wants the LaTeX edited directly: "把这段改成...", "polish this paragraph", "把我对 intro 的想法写成 LaTeX", "tighten this". No review panel; go straight to drafting the patch through the writing toolkit, with author sign-off.
• Review mode (occasional, pre-submission). The user wants the paper critiqued or hardened: review / critique / 审稿 / 评审 / mock-review, or iterating a draft to clear reviewer-raised issues. This runs the courtroom review engine (references/review-engine-v3.md).
• Auto mode (unattended). The user opts in via /goal (or config mode: auto) to run the review-revise loop AFK toward a verifiable goal. Establish the spine up front (the one human step), then the engine applies safe fixes under the bounded-aggressive policy and queues the rest. The drafter input passes the significance floor (node scripts/ledger.js floor: valid-fixable majors only) and the ledger view is initialized collapsed (--display collapse: minors fold into a Minor digest, majors stay itemized). See references/auto-mode.md. Never self-detect auto; it is explicit only.

Do NOT use for: writing a paper from scratch (use ml-paper-writing), figure or diagram generation (use academic-plotting), or an official-venue rebuttal (this is a pre-submission self-hardening loop, no score gate).

Soft update reminder: at the start of each PaperJury invocation, before choosing the mode or editing a manuscript, run node scripts/check-update.js from the skill root unless PAPERJURY_DISABLE_UPDATE_CHECK=1 is set. If it reports an available update, show the notice once and continue. If the check is skipped, silent, or cannot reach GitHub, continue without mentioning it; update checks are never allowed to block review or editing.

The three primitives

This paradigm is expressed as Skill + Workflow + Memory. Each carries one concern; together they replace the heavy per-round file-and-flag machinery a hand-rolled version accumulates.

1. Skill (this folder) = entry point + methodology. The protocol, the reviewer panel, the contestability routing, the writing toolkit, the human gates. Detail in references/review-engine-v3.md, references/reviewer-personas.md, references/writing-toolkit.md.
2. Workflow = fan-out engine. The semantic, no-human-in-the-middle steps run as Workflows (parallelism + schema-validated output by construction). The simple panel is workflows/review-panel.workflow.js; the v3 courtroom engine is assign-reviewers -> reading-check -> coverage-auditor -> merge -> {trial (+ escalate) || polish} -> recall-audit -> drafter -> {edit-audit | meaning-audit} -> clerk. The DETERMINISTIC guards run orchestrator-side via Bash between workflow calls (the Workflow sandbox has no fs): scripts/ holds decompose, extract-docx, ledger, journal, apply-patch, anchor-diff, cross-ref, spine, rekey, compile-guard, compliance-check (plus doctor, the install/repo health check: npm run doctor). Build note: this harness delivers a workflow's args as a JSON STRING, so every workflow parses it defensively. Protocol + every orchestrator seam: references/review-engine-v3.md.
3. Memory = durable state + learned conventions. Two layers:
   • Ledger (LEDGER.json resolved at runtime = the machine source of truth, plus a rendered LEDGER.md view; managed by scripts/ledger.js): the live, mutable issue state across rounds and sessions. Schema + status state machine: references/ledger-schema.md.
   • Claude memory (the active project's memory): stable conventions worth recalling next session, e.g. this paper's house style, venue, persona tuning.

Resolving inputs at runtime (no hardcoded paths)

The skill ships ZERO hardcoded paths or project files. On trigger it resolves each input by discovery first, then asking:

• manuscript: detect the main source, then route it through the INTAKE FORMAT GATE by extension. Four routes, none silent:

  • .tex: the native LaTeX path. Detect the main source (the .tex with \documentclass / \begin{document}, or the file the user names). If several candidates, ask.
  • .md / .markdown / .txt: the native text path. The full multi-round engine runs; compile checks are not applicable (compile-guard returns compiled:null plus a markdown sanity lint, an honest UNKNOWN, never a fake pass); LaTeX-only compliance checks are skipped and reported as skipped_checks.
  • .docx: if a .paper-review/ working copy AND a ledger already exist, REUSE them, never re-extract. If the sha256 of the docx no longer matches the ledger's meta.original_sha256, STOP and ask: continue on the working copy, or extract --force knowingly discarding the applied edits (an explicit new-intake event). Otherwise run node scripts/extract-docx.js extract <file.docx> (one time) and tell the user explicitly: the original Word file is never modified; all rounds run on .paper-review/<basename>.md (print the full working-copy path); they get back the edited Markdown plus a per-edit change list; the extraction report lists everything dropped or degraded. Write ledger meta {manuscript: <working copy>, working_format: 'markdown', source_format: 'docx', original, original_sha256, extracted_at, extraction_report}. If the report shows nonzero tracked-change counts, seed a round-1 author-required ledger row ("manuscript contains unresolved tracked changes; accepted-all for review").
  • any other extension (.doc, .pdf, .rtf, .odt, ...): explicitly unsupported. Say so and suggest exporting .docx / .md / .tex; never silently degrade.

  After intake, the working copy IS the manuscript for every rule and gate in this file (sign-off, spine freeze, round-0 baseline, edit safety, journal); the original uploaded file is permanently read-only.

• venue_family: the user can name it, or an agent reads the class file to GUESS the family (e.g. a cvpr/iccv style, an acl style, a neurips/iclr style). There is no hardcoded venue list and no deterministic detector; if unclear, ask.

• ledger: default to <manuscript-dir>/.paper-review/LEDGER.json (the machine source of truth; scripts/ledger.js also renders a LEDGER.md view). Create if absent, reuse if present. The user may point elsewhere.

• author: ask who signs off on edits (default: the current user). Every edit needs explicit authorization.

• personas: default to N domain-expert holistic reviewers assigned at runtime (assign-reviewers, from the project gatekeeper core + a generated domain overlay); the three generic lenses in references/reviewer-personas.md are the degrade fallback. If the project defines its own named reviewer subagents, use them as agentType; otherwise inline the persona prompts.

• style_profile: start from the venue-family default; refine from any conventions recalled from memory or pinned in a project config.

A project MAY pin these by dropping a config in ITS OWN repo (see configs/config-template.md for the shape). That file is owned by the project, never by this skill. At round start, recall any pinned conventions from memory.

Direct-edit mode (the common case)

The user states a change in Chinese or English; you draft and apply the LaTeX edit. No panel, no ledger, no discussion. Minimal flow:

1. Locate. Resolve the manuscript and find the target passage the instruction refers to (a paragraph, sentence, caption, table cell). If it is ambiguous on a large file, ask which passage; do not guess. On a .docx: if a working copy already exists, it IS the manuscript, edit it; if none exists, offer an explicit choice between (a) paste-back, returning the rewritten passage as text for the user to apply in Word (no working copy), and (b) running the one-time intake extraction and editing the working copy. Never edit the .docx file itself.
2. Draft. Pick the writing-toolkit prompt matching the instruction (translate-to-english for a Chinese idea, polish-english / de-ai for a rewrite, compress / expand for length, caption / experiment-analysis for those units) and draft the patch to do exactly what was asked. The Common guards apply (markup-safe for the working format, plain CS prose, no log leakage into the manuscript).
3. Self-gate. Run logic-check on the drafted passage.
4. Sign-off. Show the patch and get explicit author approval (hard rule 1).
5. Apply. Write only the patch into the manuscript; keep any back-translation or note author-side.

This is the writing toolkit used on its own. Escalate to review mode only when the user wants the paper critiqued or hardened, not for a single asked-for edit.

Why fan-out is a Workflow and the rest is conversation

The reviewer panel and the trial jury are pure fan-out: spawn, collect, merge. A Workflow does this deterministically (parallelism enforced by construction, structured outputs via schema, isolation by default since each agent sees only the prompt you give it). That isolation is what replaces the snapshot-and-whitelist defense: a reviewer cannot see peers, the ledger, or prior rounds because you simply do not put them in its prompt.

But the loop has genuine human gates (the author reviews the issue list, gives per-issue direction, authorizes edits, breaks ties). Workflows run to completion and return a result; they do not pause mid-run for hours of human input. So:

• fan-out steps (reviewers, trial, polish, recall, merge) -> Workflow
• human gates (per-issue direction, authorization, override) -> main conversation turns
• cross-round truth (the ledger) + stable conventions -> Memory

Review mode: one round, end to end

The full adversarial loop (the v3 courtroom engine). Use it to harden the paper, not for a single asked-for edit (that is direct-edit mode). Full protocol + the 14 orchestrator seams: references/review-engine-v3.md. [WF] = Workflow step, [det] = deterministic Node guard run orchestrator-side between workflow calls, [HUMAN] = author gate, [LEDGER] = state write.

1. Resolve + recall. Resolve the inputs above; recall this paper's conventions from memory. Pick scope: full (whole paper) or passage (one section / para / claim).
2. [det] decompose. Split the manuscript into reading units + stable passage_ids + the canonical section list.
3. [WF] assign-reviewers + [HUMAN] confirm. Name N subfields (2-4, default 3); instantiate N holistic domain reviewers from the gatekeeper core + a generated overlay. An unconfirmable slot degrades per slot to a generic gatekeeper (the three generic lenses in reviewer-personas.md are the fallback). The author confirms the assignment (or pins it via config).
4. [WF] reading-check. Each reviewer reads the WHOLE paper → weaknesses {significance(major|minor), kind(mechanical|substantive), verbatim quote — cannot quote = did not read} + one overall_confidence + a per-section coverage report. Anti-skim is three layers: [det] per-section quote-verify, [WF] coverage-auditor, [WF] targeted re-invoke.
5. [WF] merge. Semantic dedup across reviewers; derive significance (MAX) / kind (substantive-dominates) / corroboration. [LEDGER] intake as raised.
6. [det] route. mechanical → polish; substantive&minor → polish; substantive&major → trial (two parallel tracks).
7. [WF] trial. Per substantive-major charge: a whole-paper DEFENSE → 5 decorrelated local-context jurors (+ on-demand expansion) → a deterministic verdict (decide iff quorum surviving >= ceil(0.8*jurySize) AND one side > 60% of surviving votes; else escalate to 12). Verdict ∈ {invalid-drop, valid-fixable, author-required, escalate}; the judge sets a close_criterion ONLY for a valid-fixable charge, satisfiable by editing existing text (no new data). [WF] polish runs the off-gate mechanical/minor track in parallel (never silently dropped).
8. [WF] recall-audit. Mode A revives wrongly-dropped charges; Mode B spot-checks strong-consensus majors BEFORE the edit. Runs before the drafter.
9. [HUMAN] Authorize + [WF] drafter + edit-safety. On authorization, the drafter writes the minimal patch per surviving valid-fixable. The edit-safety chain gates it: [det] anchor-diff + cross-ref → [WF] meaning-audit (frozen anchor, four-state) / edit-audit (risky non-anchor); [det] apply-patch + compile-guard land a passing patch and [LEDGER] mark closed; a drift / anchor / failed edit is reverted and queued. Revision logs / back-translations stay author-side.
10. [WF] clerk + report. The clerk reconciles the round boundary (carried open-questions vs this round's edits, via a passage_id + similarity merge key) and emits convergence counts. Summarize new/closed counts with the minor/polish part as a one-line digest (counts), never per-item paragraphs; in review mode do not auto-start the next round (auto mode drives the outer loop via /goal). The rendered LEDGER.md obeys meta.display_mode (flip anytime: node scripts/ledger.js mode <ledger.json> <show|collapse>; review defaults to the flat table, auto initializes collapsed). At round end run node scripts/rekey.js <working file> <ledger> <journal> to re-link open rows whose passage_id no longer resolves after this round's edits (both formats).

GATE: node scripts/ledger.js gate = 0 gate-blocking active major (gate-blocking = {raised, in-trial, re-trial, valid-fixable}; author-required / queued / dropped / closed are gate-OK and author-required accumulates to the queue). Full protocol + ledger schema + status machine: references/review-engine-v3.md, references/ledger-schema.md. The legacy single-pass 3-reviewer panel (workflows/review-panel.workflow.js, the discussion-mode flow in references/methodology.md) is kept only as a quick check.

Hard rules (load-bearing, venue-agnostic)

1. Never edit the manuscript without explicit author sign-off. Auto-mode carve-out: the rule HOLDS; auto satisfies it via UP-FRONT sign-off (the spine confirmation + the pre-authorized bounded-aggressive policy) plus the return queue, not per-edit sign-off. Nothing outside the authorized envelope is applied.
2. Reviewers / jurors are isolated. Fresh eyes per round: no cross-talk, no prior-round leakage, no sight of the ledger. Enforced by (a) what goes into each agent's prompt AND (b) an explicit ISOLATION instruction in every reviewer-type prompt telling the agent to judge only the quoted text and not read files (workflow agents have read tools and will otherwise sometimes roam).
3. A valid-fixable issue carries a close_criterion (one concrete sentence an edit must satisfy), set by the judge at trial; it is null at intake.
4. No leakage into the reviewed text. Revision logs, back-translations, and self-check verdicts are author-side aids; they never enter the manuscript or any frozen snapshot.
5. Disagreement resolves through discussion, then override (logged), never a silent dismissal.
6. No hardcoded paths or project files in the skill. Resolve at runtime.

Memory convention

• At round start: recall the paper's conventions (house style, venue, persona tuning) from memory; read the resolved LEDGER.json for open issues.
• During the round: the ledger is the only mutable truth; update it at merge, trial verdicts, recall, and close.
• After the round: persist any newly learned stable convention to memory (e.g. a house-style rule a reviewer surfaced), not the transient issue state.

Maximizing it under ultracode

The fan-out engine implements the strong form directly (workflows/review-panel.workflow.js):

• loop-until-dry: re-runs independent fresh panels and accumulates only issues not seen before, stopping after dryStop consecutive passes that add no surviving issue (hard cap maxRounds). Raises recall past a single pass.
• adversarial verify: each new issue faces perspective-diverse skeptics (misreading / already-addressed / scope-or-severity) and is kept unless a majority refute it, filtering plausible-but-wrong issues before they reach the ledger. Bias is to keep, so real flaws are not lost.

Toggle via args: ultracode on -> defaults (maxRounds 4, dryStop 2, verify true); ultracode off -> pass {maxRounds:1, verify:false} for the basic single-panel form. The loop is budget-aware and stops early if the token budget runs low.

Capabilities and status

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); auto mode (the review-revise loop toward a goal under a drift-bounded policy, applying safe fixes and queueing risky ones for author review); and the significance floor (ledger.js floor gates the drafter to valid-fixable majors; the collapsed ledger view folds minors into a digest so trivia never floods the author's attention -- render-only, full detail kept in LEDGER.json). Roadmap: vision-based layout verification, automatic venue detection from the class file, and reviewer personas tuned to each venue community.

Related skills

• ml-paper-writing: from-scratch drafting, citation verification (never hallucinate citations), conference checklists. This loop borrows its sentence-level guidance for the edit-drafting step rather than duplicating it.
• academic-plotting: figure and architecture-diagram generation (out of scope here; this loop edits text and captions, not figure images).

[!IMPORTANT] 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、缺少证据、依赖作者私有知识或需要研究判断的问题，系统都会交回作者处理。

✨ News

• 🎉 RedNote（小红书）里程碑： 相关分享已达到 3 万次浏览、1.8k 次收藏。感谢大家的转发和收藏，也感谢大家将 PaperJury 推荐给更多正在撰写和修改论文的朋友。
• 📄 2026-06-15：PaperJury 论文已发布到 arXiv。 arXiv 页面：PaperJury: Due-Process Review for Bounded LaTeX Revision（arXiv:2606.16322）。论文系统介绍了「审稿 → 裁定 → 修改 → 复查」引擎：哪些任务由确定性脚本处理，哪些判断交给语义 agent；有争议的问题如何进入审议；不同风险的修改采用哪些护栏。
• 🔔 2026-06-10：v1.0.0 发布。 这是第一个稳定版，与 Codex 版 v1.0 对齐。新增软更新提醒：发现新的稳定 tag 时只提示，不打断当前工作。
• 🚀 2026-06-05：PaperJury 的 Codex 版已发布。 入口在这里：paperjury-codex。
• 🧪 Dogfood sample 已加入。 仓库提供一个紧凑的 dogfood sample：包括修改前后的 PDF，以及经过人工核对的运行报告。

📌 引用论文

如果 PaperJury 对你的研究或写作流程有帮助，可以引用这篇 arXiv 论文：

@misc{wang2026paperjurydueprocessreviewbounded,
  title={PaperJury: Due-Process Review for Bounded LaTeX Revision},
  author={Yiran Wang and Ruixuan An and Biao Wu and Wenhao Wang},
  year={2026},
  eprint={2606.16322},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2606.16322},
}

同一条目也放在 CITATION.bib。

⚡ 两条命令完成安装

在 Claude Code 中安装：

/plugin marketplace add Spark-To-Paper-Skills/paperjury
/plugin install paperjury@Spark-To-Paper-Skills

然后，在你的论文项目中直接说明需求：

请审稿，重点检查实验是否足以支持 claim。

也可以使用更日常的表达：

请把 introduction 这段改得更紧凑，但不要改变 claim。

无需记忆命令。PaperJury 会根据你的描述在 direct-edit 与 review 之间选择；auto 必须显式启用。在 direct-edit 和 review 模式下，补丁会先交给你确认；auto 模式则按事前授权策略应用安全修改，并将高风险修改交回作者。

📊 同一批论文、同一套评测口径

12 篇 held-out 论文(Vision、NLP、ML 各 4 篇)、四个 baseline、盲审专家审计(arXiv 2606.16322)。删除线标出的数值来自四个 baseline 中最强的 LLM-as-judge 循环:

F1：相对专家问题清单的 macro F1；Accv / Accr：盲审专家对终局裁定 / 路由决定的一致率；ESVR：已应用编辑中违反安全的比例（须与编辑量合读：PaperJury 每篇 13.4 处 vs judge 循环 14.3 处）；n/a 表示该方法没有这项能力。逐篇配对比较中，PaperJury 的 F1 在全部 12 篇上胜过每个会生成问题清单的 baseline。每篇成本为 2.47 小时、6.76M token（朴素生成器为 8.37 小时、31.4M token）。

三类裁定的盲审复核一致率分别为：invalid-drop 0.872、valid-fixable 0.913、author-required 0.860。消融实验中，移除某个部件会让与其职责相关的指标恶化（移除护栏链 → ESVR +0.152；移除庭审 → 裁定一致率 −0.153）。完整表格、图和逐项解读见 项目主页 和 论文。

🤔 先判断意见是否成立，再决定是否修改

PaperJury 以 Claude Code skill 的形式提供，把投稿前自查组织成一套闭环：审稿 → 裁定 → 修改 → 复查。它不会直接接受所有 AI 反馈，而是让每条意见经过完整的评审流程，最终得到三种裁定之一：

🎯 适合谁

📦 你会得到什么

它与一般写作工具的区别：

🧭 三种模式：direct-edit、review 与 auto

[!WARNING] auto 策略和跨轮运行都必须明确启用。 mode: auto 决定每个候选修改是自动应用还是进入待办队列；/goal 负责跨轮继续运行。仅打开工具权限或只发送普通 prompt，都不会启动无人值守的多轮循环。详见 docs/AGENT-GUIDE.md §3。

🧪 查看一次完整运行，再决定是否使用

一篇真实的 21 页草稿、11 个预先埋入的缺陷、一轮完整的 auto 模式评审：152 条 reviewer 意见去重后得到 55 个问题——其中 26 处安全应用、10 个交回作者、19 条驳回；修改后的稿件编译为 0 error / 0 warning。

samples/dogfood/(original_draft.pdf · revised_draft.pdf · 运行报告)

如果只想确认稿件不会先因格式问题被拒，可以说：

请执行 submission-readiness / 合规检查。

它会进行确定性格式筛查，再结合编译驱动的版面检查。

🚀 安装

Claude Code plugin

推荐使用 marketplace 方式安装：

/plugin marketplace add Spark-To-Paper-Skills/paperjury
/plugin install paperjury@Spark-To-Paper-Skills

Clone 成 skill

也可以把仓库 clone 到 Claude Code 读取 skill 的目录：

# macOS / Linux
git clone https://github.com/Spark-To-Paper-Skills/paperjury ~/.claude/skills/paperjury

# Windows (PowerShell)
git clone https://github.com/Spark-To-Paper-Skills/paperjury "$env:USERPROFILE\.claude\skills\paperjury"

也可以放在 <项目>/.claude/skills/ 目录下，仅对单个项目生效。

安装后建议进行以下检查：

• Claude Code 会通过 SKILL.md 自动发现该 skill，其名称为 paperjury。
• 需要安装 node，因为确定性检查运行在 Node 环境中。
• LaTeX 工具链是可选的；真实编译和版面检查会使用它，如果没有安装，会明确说明哪些检查无法执行。
• 在 skill 目录中运行 npm run doctor，可以检查仓库完整性、检查所需工具以及识别论文文件。
• 启动时会对 GitHub 稳定版 release tag 进行一次软更新检查；发现新版本时仅提示，不会阻塞当前工作。设置 PAPERJURY_DISABLE_UPDATE_CHECK=1 可以关闭提醒。更新后请开启新会话。

如何选择 Claude Code 版和 Codex 版