直接问 AI「我论文怎么样」，通常只会得到两种没用的答案：礼貌夸好，或者漫天挑刺。
PaperJury 把它改成一套闭环：先找问题，再裁定是否成立；能安全改的写成补丁，缺实验或证据的交给作者，不成立的意见直接驳回。Ask an AI if your paper is any good and you get a polite yes. PaperJury argues the other side: N reviewers read the whole paper, disputes go to an independent vote and a three-way verdict (fix it / needs you / no fix), and only edits you sign off on land, then it actually compiles.
不是三个孤立命令，而是三种使用意图：小改动走 direct-edit，投稿前自查走 review；只有需要多轮自动收敛时，才显式开启 auto。Three entries, one engine. review and auto run the same courtroom engine; auto just removes the human second instance and swaps each human gate for a pre-authorized policy + return queue.
三模式不是三个命令，而是三种意图：direct-edit 与 review 按用户的描述自动路由，只有 auto 需要你显式开启（/goal），系统不会自己切到 auto。The three modes are not three commands but three intents: direct-edit and review are auto-routed by what is said; auto is the only explicit switch (/goal) and is never self-detected.
下面是一条 issue 从提出到落稿或入队的路径。机械、轻微的问题不进庭审，走单独的快速 polish 流程；真正有争议的才进入裁定。Below is the order the engine processes one issue, step by step; mechanical and minor issues skip the trial and take a separate quick polish path.
切成 claim 单元，保留稳定 passage-idSplit into claim units + stable passage-id
decompose.js分派 N 个子领域，N 位领域 reviewer 各通读全文一遍assign N subfields; N domain reviewers each read the whole paper once
assign-reviewers · reading-check防略读审计 + 跨 reviewer 去重 → 按可争议性分流anti-skim audit + cross-reviewer dedup → route by contestability
coverage-auditor · merge全文辩护 / 5 人陪审带局部上下文 / 无多数升 12 → 法官三路裁定whole-paper defense / 5-juror local-context tier / escalate to 12 → judge routes three ways
trialA 救回误丢的问题 + B 落稿前抽检强共识 majorMode A revives wrong drops + Mode B spot-checks consensus majors before the edit
recall-auditvalid-fixable → 写最小补丁valid-fixable → a minimal patch
drafteranchor-diff + cross-ref → meaning / edit-audit → applyanchor-diff + cross-ref → meaning / edit-audit → apply
anchor-diff · cross-ref · apply-patch每轮结束时核对残留问题并判断是否收敛；review 由作者把关，auto 由 clerk 收尾round-boundary reconcile + converge; review = the author's gate, auto = clerk converges
clerk / queue每个 issue 单独处理：提出问题的 reviewer 只负责找问题，不参与裁定；独立评审组看双方证据后投票，judge 再给出三路裁定。下图沿用法庭隐喻，把检方、辩方、陪审团、法官和作者放到各自位置上。Each issue = one case; the defendant = the passage under attack, not the issue itself. Position is stance: prosecution left, defense right, jury center, judge above, appeal outside.
| 法庭角色Courtroom role | 引擎实体Engine entity | 立场 / 职责Stance / duty | ⚠ 别混淆⚠ Do not confuse |
|---|---|---|---|
| 指控 chargecharge | 一条 issueone issue | 受审的「案由」the matter on trial | 是「指控」，不是被告、不是 agentit is the charge, not the defendant, not an agent |
| 被告defendant | 受攻击的 passage / claimthe passage / claim under attack | 被动，不发言passive, does not speak | 不是 agent，不是 issuenot an agent, not an issue |
| 检方prosecution | N 位领域 reviewer（各通读全文）N domain reviewers (each reads the whole paper) | 提 charge + 引文files charge + quote | 提完退场，不 judge / 不为自己的 issue 辩steps away after filing, does not judge / does not defend its own issue |
| 辩方defense | author agentauthor agent | 带证据为其辩护：已处理 / 越界 / 改了会漂移argues its case with evidence: already addressed / out of scope / fixing it would drift | = author AGENT，不是人类；之后转为 drafter= author AGENT, not the human; later becomes the drafter |
| 覆盖 + 合并coverage + merge | coverage-auditor + merge(workflow)coverage-auditor + merge (workflows) | 防略读审计 + 跨 reviewer 去重 + 按可争议性分流anti-skim audit + cross-reviewer dedup + route by contestability | 机械 / 轻微件走 polish 流程，不进庭审mechanical / minor items take the polish track, not the trial |
| 庭审陪审团trial jury | 先 5 位、必要时升到 12 位、视角彼此独立的全新陪审员5, escalating to 12, fresh jurors with independent perspectives | 中立，带着局部上下文、凭双方证据投票neutral, votes on both sides' evidence with local context | fresh，既非 reviewer 也非 author；只有无明显多数才升 12fresh, neither reviewer nor author; escalates to 12 only on no clear majority |
| 法官judge | 1 presiding agent1 presiding agent | 主持、汇总、裁定、设定 criterionpresides, tallies, rules, sets criterion | 只裁定，不投票rules only, does not vote |
| 起草 drafterdrafter | author agent（转换角色）author agent (different stage) | 对 valid-fixable 起草 minimal-edit；需新数据则诚实软化或升人minimal-edit for valid-fixable; for needs-data, honest softening or escalate | 同 persona、不同阶段same persona, different stage |
| 召回审计recall audit | fresh skepticfresh skeptic | A 复查每条 drop；B 落稿前抽检强共识 majorMode A re-checks every drop; Mode B spot-checks consensus majors before the edit | 不是原 reviewer（避免固守先前判断）not the original reviewer (avoids entrenched positions) |
| 二审 / 终审second / final | 人类作者the human author | 只看升级件，终裁记录在案sees only escalations, final ruling logged | = 人，不是 author agent= the human, not the author agent |
| 执行官 / orchestratorbailiff / orchestrator | orchestrator + 确定性脚本orchestrator + deterministic scripts | 跑护栏 / 维护 ledger / 写 journalruns guards / keeps ledger / writes journal | 庭外，不 judge，只跑确定性 + gateoff-court, does not judge, only runs the deterministic guards + gate |
| 书记官 clerkclerk | clerk(workflow)clerk (workflow) | 轮边界：核对遗留问题 vs 本轮改动、去重、判定是否收敛round boundary: reconcile carried open-questions vs this round's edits, dedup, decide convergence | 语义 workflow，不是上面的法警（确定性脚本）a semantic workflow, not the bailiff (deterministic scripts) above |
分工原则很简单：找全靠阅读，判准靠裁定。阅读阶段只负责找问题和给引文，不负责判定是否成立；每条观察都必须引用原文（引不出 = 没真读，也可能是在编）。N 位领域 reviewer 各自通读全文一遍，再用三层机制防略读。Organizing principle: recall belongs to reading, precision to the courtroom. Readers only find + cite, they do not judge validity; they must quote the source (cannot quote = did not read, anti-skim and anti-hallucination). N domain reviewers each read the whole paper once, backstopped by a three-layer anti-skim.
PaperJury 不默认相信 reviewer 意见。每个 issue 先核验，再按性质分流：能安全改的写补丁；缺实验、缺证据或需要研究判断的交给作者；不成立的直接驳回。Routes in two stages: first a deterministic contestability router sends mechanical + minor issues to a separate polish track (no jury); only substantive-major issues reach the trial. The three routes below are the judge's verdict inside the trial, downstream of the contestability router.
误读、幻觉、重复、越界或严重度虚高：驳回，并保留证据。accurately judged invalid (hallucination / misread / duplicate / out of scope / inflated severity) → dropped, with evidence.
→ 仍会交给召回审计抽查→ to recall audit for re-check文本问题、claim 过强或结构不清；无需新实验或数据，也不会漂移：写最小补丁，过护栏后落稿。valid + text-fixable (no new experiments/data) + no drift → draft a minimal edit, pass the guards, land it.
→ drafter · 需要新数据则改判 author-required 或诚实软化→ drafter · if it needs new data, reroute to author-required or soften honestly缺实验、缺数据，或需要研究判断：进入作者待办队列。这是正确分流，不算漏判。needs author-private info / new experiments / a judgment call → routed to the author (second instance). Correct handling, not a recall loss.
→ review 交作者 · auto 入队→ review hands it to the author · auto queues it