https://cygnusyang.github.io/posts/01-ai-%E5%B7%A5%E5%85%B7%E4%B8%8E%E6%99%BA%E8%83%BD%E4%BD%93/08-academic-research-skills/07-pipeline/技术 × 产品 × 机器人Hugo 0.160.1 & FixIt v0.4.6-20260517062719-ed8ec5c2zh-CNruohuyang@163.com (Cygnus Yang)ruohuyang@163.com (Cygnus Yang)Sat, 23 May 2026 00:00:00 +0800https://cygnusyang.github.io/posts/01-ai-%E5%B7%A5%E5%85%B7%E4%B8%8E%E6%99%BA%E8%83%BD%E4%BD%93/08-academic-research-skills/07-pipeline/14-10-stage-orchestration/Sat, 23 May 2026 00:00:00 +0800ruohuyang@163.com (Cygnus Yang)https://cygnusyang.github.io/posts/01-ai-%E5%B7%A5%E5%85%B7%E4%B8%8E%E6%99%BA%E8%83%BD%E4%BD%93/08-academic-research-skills/07-pipeline/14-10-stage-orchestration/<p>Academic Pipeline 是 ARS 的&quot;总调度&quot;——它将 Deep Research、Academic Paper 和 Academic Paper Reviewer 三个独立 Skill 编排为端到端的学术生产管线。但它不是简单的顺序调用器。Pipeline 在三个阶段之间插入了两次诚信闸门、一次决策分支、两次苏格拉底式指导环节和一个协作评估层。</p> <p>本文从编排器的视角，详解 10 个阶段的流转逻辑、checkpoint 机制和决策分支。</p> <h2 class="heading-element" id="10-阶段全览"><span>10 阶段全览</span> <a href="#10-%e9%98%b6%e6%ae%b5%e5%85%a8%e8%a7%88" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><pre><code>Stage 1: RESEARCH 🧑 用户确认 RQ &#43; 方法论 Stage 2: WRITE 🧑 用户批准大纲 Stage 2.5: INTEGRITY ✓ 7 类失败模式检查 Stage 3: REVIEW 🧑 用户做出编辑决定 ├── Accept ──────────────→ Stage 4.5 ├── Minor/Major → Coaching → Stage 4 └── Reject ──────────────→ 结束 Stage 4: REVISE 🧑 用户确认修订 Stage 3&#39;: RE-REVIEW 🧑 用户确认复审决定 Stage 4&#39;: RE-REVISE 🧑 内容冻结 Stage 4.5: FINAL INTEGRITY ✓ 零容忍最终验证 Stage 5: FINALIZE 🧑 用户选择输出格式 Stage 6: PROCESS SUMMARY 🧑 最终确认</code></pre><p>🧑 = 决策重点 | ✓ = 诚信闸门</p>https://cygnusyang.github.io/posts/01-ai-%E5%B7%A5%E5%85%B7%E4%B8%8E%E6%99%BA%E8%83%BD%E4%BD%93/08-academic-research-skills/07-pipeline/15-integrity-gates/Sat, 23 May 2026 00:00:00 +0800ruohuyang@163.com (Cygnus Yang)https://cygnusyang.github.io/posts/01-ai-%E5%B7%A5%E5%85%B7%E4%B8%8E%E6%99%BA%E8%83%BD%E4%BD%93/08-academic-research-skills/07-pipeline/15-integrity-gates/<p>诚信闸门是 ARS pipeline 中最不可妥协的设计。两个闸门在代码层面被硬编码为不可跳过——与那些&quot;用户可以点继续跳过&quot;的 checkpoint 不同，Stage 2.5 和 Stage 4.5 的检查结果必须被用户逐条审阅并明确确认。</p> <p>这不是&quot;信任 AI 的判断&quot;，而是&quot;将 AI 的判断结构化呈现给人类&quot;。</p> <p>本文详解两个闸门的检查内容、运行机制、阻断逻辑和 v3.4.0 Compliance Agent 的整合。</p> <h2 class="heading-element" id="两个闸门的分工"><span>两个闸门的分工</span> <a href="#%e4%b8%a4%e4%b8%aa%e9%97%b8%e9%97%a8%e7%9a%84%e5%88%86%e5%b7%a5" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><table> <thead> <tr> <th>属性</th> <th>Stage 2.5</th> <th>Stage 4.5</th> </tr> </thead> <tbody> <tr> <td>时机</td> <td>论文初稿完成后、进入同行评议前</td> <td>所有修订完成后、格式转换前</td> </tr> <tr> <td>目标</td> <td>在进入审稿前发现并解决诚信问题</td> <td>在出版前以零容忍标准做最终验证</td> </tr> <tr> <td>声明验证范围</td> <td>30% 抽样（最少 10 条）</td> <td>100% 全部声明</td> </tr> <tr> <td>重试机制</td> <td>最多 3 次</td> <td>无重试——FAIL 阻断</td> </tr> <tr> <td>历史追踪</td> <td>首次检查</td> <td>对比 2.5 的结果，任何 SUSPECTED → 必须 CLEAR 或 Override</td> </tr> </tbody> </table> <hr> <h2 class="heading-element" id="stage-25-的检查内容"><span>Stage 2.5 的检查内容</span> <a href="#stage-25-%e7%9a%84%e6%a3%80%e6%9f%a5%e5%86%85%e5%ae%b9" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><h3 class="heading-element" id="1-7-类-ai-研究失败模式检查"><span>1. 7 类 AI 研究失败模式检查</span> <a href="#1-7-%e7%b1%bb-ai-%e7%a0%94%e7%a9%b6%e5%a4%b1%e8%b4%a5%e6%a8%a1%e5%bc%8f%e6%a3%80%e6%9f%a5" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h3><p>Integrity Verification Agent 对论文逐条运行在<a href="../../01-philosophy/02-ai-research-failure-modes.md">第二篇文章</a>中详细讨论过的 7 类检测问题。每条模式输出三个可能结论：</p>https://cygnusyang.github.io/posts/01-ai-%E5%B7%A5%E5%85%B7%E4%B8%8E%E6%99%BA%E8%83%BD%E4%BD%93/08-academic-research-skills/07-pipeline/16-material-passport/Sat, 23 May 2026 00:00:00 +0800ruohuyang@163.com (Cygnus Yang)https://cygnusyang.github.io/posts/01-ai-%E5%B7%A5%E5%85%B7%E4%B8%8E%E6%99%BA%E8%83%BD%E4%BD%93/08-academic-research-skills/07-pipeline/16-material-passport/<p>学术研究不是一次性的对话。一篇论文从构思到投稿可能跨越数天甚至数周，中间经过多次&quot;暂停—思考—继续&quot;的循环。ARS 的 Material Passport（Schema 9）就是为这种工作模式设计的——它是一份随 pipeline 流转的 YAML 文档，记录所有阶段的产出物哈希、合规历史和 checkpoint 决策，让你可以随时中断和恢复研究。</p> <p>本文详解 Material Passport 的数据结构、输入端口（literature_corpus[]）、重置边界机制和可复现性文档（repro_lock）。</p> <h2 class="heading-element" id="passport-的设计哲学"><span>Passport 的设计哲学</span> <a href="#passport-%e7%9a%84%e8%ae%be%e8%ae%a1%e5%93%b2%e5%ad%a6" class="heading-mark"> <svg class="octicon octicon-link" viewBox="0 0 16 16" version="1.1" width="16" height="16" aria-hidden="true"><path d="m7.775 3.275 1.25-1.25a3.5 3.5 0 1 1 4.95 4.95l-2.5 2.5a3.5 3.5 0 0 1-4.95 0 .751.751 0 0 1 .018-1.042.751.751 0 0 1 1.042-.018 1.998 1.998 0 0 0 2.83 0l2.5-2.5a2.002 2.002 0 0 0-2.83-2.83l-1.25 1.25a.751.751 0 0 1-1.042-.018.751.751 0 0 1-.018-1.042Zm-4.69 9.64a1.998 1.998 0 0 0 2.83 0l1.25-1.25a.751.751 0 0 1 1.042.018.751.751 0 0 1 .018 1.042l-1.25 1.25a3.5 3.5 0 1 1-4.95-4.95l2.5-2.5a3.5 3.5 0 0 1 4.95 0 .751.751 0 0 1-.018 1.042.751.751 0 0 1-1.042.018 1.998 1.998 0 0 0-2.83 0l-2.5 2.5a1.998 1.998 0 0 0 0 2.83Z"></path></svg> </a> </h2><p>Material Passport 解决的核心问题是：<strong>跨 session 状态管理</strong>。</p>