{
  "schema_version": 3,
  "paper": {
    "title": "CoworkOS: Scaling Agent Coworkers for Real-world Automation",
    "kind": "research",
    "primary_source_ids": [
      "paper"
    ],
    "arxiv_id": "2611.03178",
    "version": "v1",
    "earliest_public_date": "2026-11-02"
  },
  "artifacts": {
    "article": "content/_posts/099-coworkos.md",
    "deck": "lib/decks/coworkos.html",
    "embed_src": "/lib/decks/coworkos.html"
  },
  "sources": [
    {
      "id": "paper",
      "kind": "paper",
      "title": "CoworkOS: Scaling Agent Coworkers for Real-world Automation",
      "url": "https://arxiv.org/pdf/2611.03178v1",
      "revision": "v1",
      "used_for": [
        "problem",
        "method",
        "results",
        "limitations"
      ]
    },
    {
      "id": "metadata",
      "kind": "paper",
      "title": "arXiv v1 date and CC BY-NC-SA 4.0 license",
      "url": "https://arxiv.org/abs/2611.03178",
      "revision": "v1",
      "used_for": [
        "date",
        "title",
        "license"
      ]
    },
    {
      "id": "code",
      "kind": "code",
      "title": "Official CoworkOS harness implementation",
      "url": "https://github.com/RUC-NLPIR/CoworkOS",
      "revision": "baf03c30db56160690f9d03bfa7dee925bad6cfd",
      "used_for": [
        "coordination",
        "resource-revision",
        "persistence",
        "context",
        "release"
      ]
    },
    {
      "id": "project",
      "kind": "website",
      "title": "Official gallery 24 cases",
      "url": "https://coworkos-ai.github.io/showcase.html",
      "checked_on": "2026-11-09",
      "used_for": [
        "release",
        "qualitative-scope"
      ]
    },
    {
      "id": "review",
      "kind": "website",
      "title": "Author agent review protocol and Ridge Pack action records",
      "url": "https://coworkos-ai.github.io/review/web/coverage.json",
      "checked_on": "2026-11-09",
      "used_for": [
        "real-case",
        "verification-boundary"
      ]
    }
  ],
  "paper_reading": {
    "main_agent_complete_read": true,
    "appendices_included": true,
    "extracted_text_reviewed": true,
    "rendered_pages_reviewed": true,
    "page_count": 34,
    "rendered_page_ranges": [
      {
        "start": 1,
        "end": 34
      }
    ],
    "source_visuals_expected": 17,
    "visual_inventory_complete": true,
    "pdf_sha256": "0db3eb94695af5b2bd6c6da600691091202d5e0df79d850ca734f2a85493cf15",
    "text_paths": [
      "work/paper-2611-03178/paper.txt"
    ],
    "rendered_path": "work/paper-2611-03178/page-01.png through page-34.png",
    "review_scope": "Full main-agent read resumed across compaction; original crops individually inspected; no independent experiment reproduction."
  },
  "note_sections": [
    {
      "role": "problem",
      "locator": "1. Motivation：一条最终回复，接不住持续变化的工作"
    },
    {
      "role": "related-work",
      "locator": "2. Related Work：从编码会话扩展到共同工作的运行环境"
    },
    {
      "role": "method",
      "locator": "3. Method：谁做决定，谁保存状态，谁提交修改"
    },
    {
      "role": "experiments",
      "locator": "4. Experiments：框架比较、真实产物和五轮SFT分别证明什么"
    },
    {
      "role": "future-work",
      "locator": "5. Discussion：有价值的系统设计，还缺哪些可验证证据"
    },
    {
      "role": "summary",
      "locator": "6. Conclusion：运行时连续性值得看，训练收益要按SFT理解"
    }
  ],
  "terminology": [
    {
      "term": "harness",
      "kind": "operation",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§3.1; Part0"
        }
      ],
      "article_quote": "模型负责读上下文、提出操作和解释结果；运行框架（harness）负责实际执行工具、构造下一次上下文、保存状态及安排等待。交付物由具体任务决定，可以是可交互网页、文档、CAD工程或模型训练结果，并没有统一要求输出一个index.html。本文的Main Agent是直接承接用户目标的主执行者；Coordinator Runtime是管理状态、消息与执行生命周期的程序，不是另一位自主制定研究策略的模型。"
    },
    {
      "term": "Main Agent",
      "kind": "actor",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§3.1.1"
        }
      ],
      "article_quote": "模型负责读上下文、提出操作和解释结果；运行框架（harness）负责实际执行工具、构造下一次上下文、保存状态及安排等待。交付物由具体任务决定，可以是可交互网页、文档、CAD工程或模型训练结果，并没有统一要求输出一个index.html。本文的Main Agent是直接承接用户目标的主执行者；Coordinator Runtime是管理状态、消息与执行生命周期的程序，不是另一位自主制定研究策略的模型。"
    },
    {
      "term": "Coordinator Runtime",
      "kind": "actor",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§3.1.1"
        }
      ],
      "article_quote": "模型负责读上下文、提出操作和解释结果；运行框架（harness）负责实际执行工具、构造下一次上下文、保存状态及安排等待。交付物由具体任务决定，可以是可交互网页、文档、CAD工程或模型训练结果，并没有统一要求输出一个index.html。本文的Main Agent是直接承接用户目标的主执行者；Coordinator Runtime是管理状态、消息与执行生命周期的程序，不是另一位自主制定研究策略的模型。"
    },
    {
      "term": "Subagents",
      "kind": "actor",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§3.1.1"
        }
      ],
      "article_quote": "Main Agent直接接收用户请求，选择自己执行还是委派，并决定最后是否满足目标。可选的Subagents是接到具体子任务的独立执行者；它们获得有限背景、任务说明、明确的资源权限和技能目录，不会自动拿到Main的全部历史和直接用户输入。Coordinator Runtime负责持久任务记录、排队、生命周期和消息投递，执行Main提出的安排。"
    },
    {
      "term": "Inbox",
      "kind": "state",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§3.1.2"
        }
      ],
      "article_quote": "同步委派沿调用路径等待结果；异步委派先返回确认和标识，Main可以继续别的工作。子任务之后把结果送入持久收件队列（Inbox），运行时在安全的模型决策边界将消息交给Main。它不会把后台结果任意插入尚未结束的工具调用中。子Agent之间可以交换消息，但这不产生第二套目标控制层。"
    },
    {
      "term": "baseRevision",
      "kind": "state",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§3.2.1 Eq.7"
        }
      ],
      "article_quote": "资源协议用identity、kind、content和revision描述共享对象：标识指向同一对象，类型决定应用能力，内容是当前产物，版本号用于发现并发变化。提交修改时携带读入时的baseRevision。Eq.7要求baseRevision等于currentRevision才能应用；不一致则报告冲突，保留候选并重新考虑。"
    },
    {
      "term": "Job",
      "kind": "operation",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§3.3.1"
        }
      ],
      "article_quote": "后台进程（Job）把命令、日志、进程身份和退出结果单独保存，使训练或服务可以在Main换上下文、重新启动之后继续被跟踪。主机进程仍存在是关键条件；主机消失或外部服务失败时，不能无条件承诺恢复成功。当前jobs.ts在进程已消失但没有退出记录时报告uncertain/unobserved，不把消失误当成功。"
    },
    {
      "term": "parking",
      "kind": "operation",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§3.3.1"
        }
      ],
      "article_quote": "监控源的轮询频率与Main审阅目标的频率是两回事。确定性监控只检查状态，未变化时可以不调用模型。暂时停下来等待（parking）必须有计时器、活动监控源，或者目标自身的下一次审阅条件；公开sleep_until在没有唤醒条件时返回parked:false和no_wake_condition。任务结束前，运行时还检查未结束Job、子任务、回调和未处理用户请求；产物质量是否合格仍由Main判断。"
    },
    {
      "term": "SFT",
      "kind": "loss",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§4.3.2; Appendix A.3"
        }
      ],
      "article_quote": "CoworkOS研究怎样让模型在复杂工作中持续执行、等待反馈，并与人一起修改同一份产物。它提供模型周围的运行框架：安排工具与子任务，保存工作状态，处理后台结果和共享文件。论文另用这套框架组织五轮监督微调实验；这一部分训练的是Qwen模型，采用SFT，没有报告RL。"
    }
  ],
  "coverage": [
    {
      "id": "central-claim",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§1; §3"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "CoworkOS研究怎样让模型在复杂工作中持续执行、等待反馈，并与人一起修改同一份产物。它提供模型周围的运行框架：安排工具与子任务，保存工作状态，处理后台结果和共享文件。论文另用这套框架组织五轮监督微调实验；这一部分训练的是Qwen模型，采用SFT，没有报告RL。"
        }
      ]
    },
    {
      "id": "motivation",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§1; Figure2"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "论文关注两种工作。第一种是多步骤项目：产物、外部服务和训练任务可能持续数小时，结果迟于模型当前回合返回。第二种是持续监督：定期观察数据或状态，发现变化后再采取行动。人还可能在过程中修改文件或工程，Agent必须知道眼前版本已经改变，而不能继续按照旧上下文覆盖它。"
        }
      ]
    },
    {
      "id": "related-work-comparison",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§2; §4.1"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "最直接的比较对象是OpenHands、Codex和DeepSeek Harness（DSH），它们同样把模型接入工具与执行环境。论文的区别在于将人和Agent共同修改资源、跨上下文持续执行、异步反馈与委派放进同一个工作区。相比只提供多个Agent聊天，它同时管理产物、进程、任务状态和验收责任。"
        }
      ]
    },
    {
      "id": "same-model-harness",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§4.1.1"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "论文主表把四个框架都接到GLM-5.3，使用High thinking effort、256K上下文和每次响应最多64K输出。这里Codex是框架名称，不能读成OpenAI模型与GLM的模型能力竞赛。同一模型减少了一类干扰，但工具实现、提示、任务适配和实际用量仍可能不同；主表没有充分的组件消融和等成本分析。"
        }
      ]
    },
    {
      "id": "main-subagent-context",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§3.1.1"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "Main Agent直接接收用户请求，选择自己执行还是委派，并决定最后是否满足目标。可选的Subagents是接到具体子任务的独立执行者；它们获得有限背景、任务说明、明确的资源权限和技能目录，不会自动拿到Main的全部历史和直接用户输入。Coordinator Runtime负责持久任务记录、排队、生命周期和消息投递，执行Main提出的安排。"
        }
      ]
    },
    {
      "id": "async-boundary",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§3.1.2"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "同步委派沿调用路径等待结果；异步委派先返回确认和标识，Main可以继续别的工作。子任务之后把结果送入持久收件队列（Inbox），运行时在安全的模型决策边界将消息交给Main。它不会把后台结果任意插入尚未结束的工具调用中。子Agent之间可以交换消息，但这不产生第二套目标控制层。"
        }
      ]
    },
    {
      "id": "implementation-inbox",
      "claim_type": "inspected-implementation",
      "source_ids": [
        "code"
      ],
      "source_locators": [
        {
          "source_id": "code",
          "locator": "coordinator-runtime.ts:657–753,1022–1034; scheduler.ts:271–328; main-inbox-delivery.ts:273–333"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "当前代码可以验证这一安排：coordination.ts创建受限委派，coordinator-runtime.ts先记录队列与checkpoint，再确认异步调用；完成记录保存后产生Main回调。scheduler.ts按messageId去重并租用消息，Main投递路径在消息进入保存的上下文后才确认消费。这里的checkpoint是运行记录，不是SFT模型权重；这些源码证据也不等于完成了进程崩溃或重复外部写入的故障测试。"
        }
      ]
    },
    {
      "id": "resource-protocol",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§3.2.1 Eq7"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "资源协议用identity、kind、content和revision描述共享对象：标识指向同一对象，类型决定应用能力，内容是当前产物，版本号用于发现并发变化。提交修改时携带读入时的baseRevision。Eq.7要求baseRevision等于currentRevision才能应用；不一致则报告冲突，保留候选并重新考虑。"
        }
      ]
    },
    {
      "id": "cad-example",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "Figure4"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "Figure 4的房屋是论文示意案例：v1是一层，Agent加二层与阳台得到v2；人把上层后退形成露台，保存成v3；Agent观察变化后补遮阴、楼梯和栏杆形成v4。这幅图解释预期协作过程，并没有公开证明每步发生过的事件日志。"
        }
      ]
    },
    {
      "id": "worked-conflict",
      "claim_type": "reproduced-calculation",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "Eq7; Fig4"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "用版本号算一次冲突：Agent的候选基于v2，用户已经保存v3，则2≠3，提交失败；候选仍保留。Main重新读v3并调整方案，新的候选以3为baseRevision，当前仍为3时才允许保存。版本一致只能证明允许提交，不能证明遮阴和楼梯设计满足人的意图。"
        }
      ]
    },
    {
      "id": "resource-code-mcp-boundary",
      "claim_type": "inspected-implementation",
      "source_ids": [
        "code"
      ],
      "source_locators": [
        {
          "source_id": "code",
          "locator": "resource-store.ts:424–474; agent-capabilities.ts:135–183,391–452; docs/external-applications.md:17"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "公开resource-store.ts确实在写锁中检查操作身份和baseRevision，过时版本抛出revision_conflict；子Agent可在私有draft中修改，再保存带版本的candidate，由Main审阅接纳。主工作区因此不必随子任务试错不断变动。但外部MCP软件保有自己的保存和并发语义：官方文档明确外部Blender不使用内置编辑器版本历史。只有明确返回资源操作、接入对应协议的路径，才可能采用这些检查。"
        }
      ]
    },
    {
      "id": "persistent-jobs",
      "claim_type": "author-claim",
      "source_ids": [
        "paper",
        "code"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§3.3.1; jobs.ts:92–176"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "后台进程（Job）把命令、日志、进程身份和退出结果单独保存，使训练或服务可以在Main换上下文、重新启动之后继续被跟踪。主机进程仍存在是关键条件；主机消失或外部服务失败时，不能无条件承诺恢复成功。当前jobs.ts在进程已消失但没有退出记录时报告uncertain/unobserved，不把消失误当成功。"
        }
      ]
    },
    {
      "id": "parking-monitor-completion",
      "claim_type": "inspected-implementation",
      "source_ids": [
        "code",
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "code",
          "locator": "§3.3.1; coordination.ts:201–281; task-completion.ts:19–40"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "监控源的轮询频率与Main审阅目标的频率是两回事。确定性监控只检查状态，未变化时可以不调用模型。暂时停下来等待（parking）必须有计时器、活动监控源，或者目标自身的下一次审阅条件；公开sleep_until在没有唤醒条件时返回parked:false和no_wake_condition。任务结束前，运行时还检查未结束Job、子任务、回调和未处理用户请求；产物质量是否合格仍由Main判断。"
        }
      ]
    },
    {
      "id": "paper-memory",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§3.3.3"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "工作区记忆是Main维护的文档，记录用户要求、结论、未完成工作和下一步，不是训练到参数里的记忆。论文§3.3.3描述显式重建上下文：结合记忆、工作区状态和近期历史继续工作，避免只依赖不断增长的聊天记录。"
        }
      ]
    },
    {
      "id": "current-clear-context",
      "claim_type": "inspected-implementation",
      "source_ids": [
        "code",
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "code",
          "locator": "workspace.ts:113–137; extension.ts:1022–1033"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "当前公开实现需要单独说明：clear_context会在当前回合结束后调用ctx.compact，生成历史摘要并保留安全的近期消息；完整消息日志与memory.md继续保存在磁盘。源码提示记忆只是补充，而且可能过时，摘要必须以实际会话为依据；压缩失败则保留原上下文。不能按工具名字把它写成“清空全部历史，只靠memory.md重建”。"
        }
      ]
    },
    {
      "id": "model-results",
      "claim_type": "reproduced-calculation",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "Table1"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "Table 1覆盖工具使用、生产力、搜索和机器学习编码四类能力，报告各框架在实际配置下的分数。CoworkOS在生产力均分57.54和ML Coding均分56.26最高；工具使用55.22略低于DSH的55.31，搜索82.14略低于OpenHands的82.22。这里支持广泛竞争力，不支持每类第一，更不能从总表分离出版本协议、子Agent或记忆各自造成了多少提升。"
        }
      ]
    },
    {
      "id": "negative-harness-results",
      "claim_type": "reproduced-calculation",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "Table1"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "几个单项保留比较条件：SkillsBench 64.34、JobBench 59.88、WideSearch 82.47、MLE-Bench-Lite 76.90与MLS-Bench 35.62都是表内最高；Harvey 22.74低于OpenHands 23.75，OfficeQA 65.90低于Codex 66.30，Workspace 67.14低于DSH 68.76。ML Coding均分仅比DSH高0.12，且没有重复seed或误差条，不能把这点差距当作稳定优势。"
        }
      ]
    },
    {
      "id": "task-provenance-scoring",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "Appendix A.1"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "分数来自不同判断机制，不能把所有行当作同一种成功率。SkillsBench是79项文本任务的平均任务分，排除视觉和媒体任务；Tau3 Banking是97项任务通过率；JobBench是全部65项产物的LLM任务分均值；OfficeQA在133题和解析后的Treasury Bulletin材料上按原生规则评分。AutomationBench报告all-pass：满足该任务要求的全部检查才通过，本文没有取得它的逐项断言或本次完整原始评测日志。"
        }
      ]
    },
    {
      "id": "missing-checker-logs",
      "claim_type": "unknown",
      "source_ids": [
        "paper",
        "code"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "AppendixA.1; test/jobbench-runtime.test.js:1–37; public tracked-file checks"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "论文说明保留各评测的默认或官方规则，但没有统一列出全部模型可见提示、judge隐藏输入、版本与运行日志。公开harness的单元测试也不能补足这些正式任务检查；例如jobbench-runtime.test.js使用mock工具，是运行时测试。"
        }
      ]
    },
    {
      "id": "gdp-main-metric",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "AppendixA.1; Table1"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "主表GDPVal的1164是相对固定1000基线的Elo。设W、L、T分别为胜、负、平，Appendix A.1给出E=1000+400log₁₀[(W+T/2)/(L+T/2)]，且W+L+T=220。失去有效评分的任务也算负，不能从分母剔除。生产力均分使用的是100(W+T/2)/220，而不是将1164直接与其它百分数平均；论文未给精确W/L/T，不能从四舍五入的Elo倒推出唯一计数。"
        }
      ]
    },
    {
      "id": "gdp-sft-worked-metric",
      "claim_type": "reproduced-calculation",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "AppendixA.3"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "后面的SFT实验换了指标。Appendix A.3的GDPVal先汇总有效判分任务的实际rubric分sᵢ和最大分mᵢ，再算100Σsᵢ/Σmᵢ；有效判分任务集合V是它的分母边界。这不是Elo、胜率，也不是每题百分数的平均。用本文说明性输入算一次：两题4/5和9/15，累计指标=100×13/20=65，逐题平均=(80+60)/2=70。高满分任务对累计指标权重更大。这个例子不是作者的真实任务结果。"
        }
      ]
    },
    {
      "id": "real-paper-case",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "AppendixA.2; Figure9"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "Ridge Pack是论文与官网共同提供的真实网页产物案例：用户可以切换背包容量、加入物品、检查分隔空间并优化或撤销。Figure 9和Appendix A.2记录作者按同一顺序尝试八个同名物品后的可见反馈。CoworkOS提示尚余5.98L，却因空间碎片化不能找到不与邻居相碰的连续位置，建议Optimize或移除较大的邻居。它解释的是总容积与可用连续空间的区别。"
        }
      ]
    },
    {
      "id": "case-baseline-boundary",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "AppendixA.2; Figure9"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "Codex版本也会拒绝Picnic blanket，并建议移除较大物品或改用18L包。DSH的截图保留了未装入blanket的列表，但没有捕获瞬态拒绝提示，因而不能宣称它“不解释”。三版使用的物品尺寸不同，装入数量和容量值不能用来排名装箱算法。"
        }
      ]
    },
    {
      "id": "real-public-actions",
      "claim_type": "inspected-implementation",
      "source_ids": [
        "review"
      ],
      "source_locators": [
        {
          "source_id": "review",
          "locator": "review/web/coverage.json records[9:12]"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "作者另公开了 review/web/coverage.json 。reviewer在12L版本依次尝试Water bottle、Camera cube、Rolled jacket、Lunchbox、Field notebook、Sunglasses case、Snack pouch和Picnic blanket，并测试容量切换、Optimize、Undo与Reset。CoworkOS记录为6件装入，jacket和blanket被拒，余量7.11/12.11L；切18L再回12L保留6件，Optimize重排但仍不能装入两件，Undo恢复，Reset回零。Codex记录装入7件，blanket仍拒绝，切换容量后保留状态。这些是作者另一段公开review的数值，不能与论文截图的5.98L合成同一条执行日志。"
        }
      ]
    },
    {
      "id": "case-validation-limits",
      "claim_type": "interpretation",
      "source_ids": [
        "paper",
        "review"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "AppendixA.2; review records verification gaps"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "该case检查的是交互是否响应、反馈能否解释拒绝，以及部分撤销与重置状态是否保留；没有官方装箱最优性PASS断言，也未全面验证拖放、移除、导出和碰撞不变量。因此它可以证明交付网页具有可操作状态，不能代替程序正确性或最优性评测。Figure 9的Courtyard House同样保留基线能力：三版都支持Kitchen inspection并保留相关设置；CoworkOS把14.9m²、4.68×3.18m、17:00、剖面1.40m与材质方案B集中显示，主要证据是信息组织。"
        }
      ]
    },
    {
      "id": "qualitative-protocol",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§4.2; AppendixA.2"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "Figure 6比较网页、游戏、Office、交互图、CAD和Blender产物。作者确实进一步执行了游戏输入、失败和重试，Office检查也保留内容页，不只看封面。但是各实现的游戏分数、物品尺寸与部分内容不同，不能用截图显示数字排成功率；Office布局清楚也不证明里面的数值正确。"
        }
      ]
    },
    {
      "id": "strongest-3d-counterexample",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "AppendixA.2; CAD/GLB inspection"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "强反例来自3D检查：VESPER Speaker在CoworkOS中反复reset、展开、等待和旋转，仍未清楚展示分离后的内部结构；Saturn Atlas的Codex剖面层次更清楚。六份STEP用OCP 8.0.1.1.0成功导入，仅建立几何可读取，不证明参数化编辑、装配干涉、公差或制造性能。公开评审GLB没有animation tracks，旋转viewport不等于作品带动画。"
        }
      ]
    },
    {
      "id": "geometry-failure",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "Fig14; AppendixA.2"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "Figure 14刻意保留After the Fall中CoworkOS的公交轮胎和下部车身悬空。这说明场景空间层次、材质和照明更丰富，仍可能支撑关系错误。三版照明和材质没有统一校准，不能把视觉偏好当物理正确性；几何反例应与漂亮总览一起读。"
        }
      ]
    },
    {
      "id": "sft-not-rl",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§4.3.2; AppendixA.3"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "五轮后训练使用固定GLM-5.3作为controller和数据teacher，Qwen3.6-35B-A3B是被训练与评测的模型。每轮先评测Qwen，诊断能力缺口，修改数据生成流程，生成、修复、验证和过滤教师轨迹，然后做监督微调（SFT）并重新评测。分数用于决定下一轮训练什么，不是直接作为RL策略梯度的奖励；论文未报告RL、GRPO、OPD或教师logits/KL蒸馏，也没有充分给出token loss mask和精确优化器。"
        }
      ]
    },
    {
      "id": "training-lineage-filter",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§4.3.2–4.3.4; AppendixA.3; p34 unnumbered table"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "R0用1345条工具交互seed从base训练；R1把244条新演示加入seed，并再次从base训练，不是接着R0微调。R2才从R1 checkpoint继续，135条首次使用数据由77条先前保留但未训练的轨迹和58条新生成轨迹组成。训练历史查重与BM25任务相似度过滤是两项检查：后者阈值0.8用于降低train/eval近重复，前者检查父模型是否已经用过数据。"
        }
      ]
    },
    {
      "id": "r2-failure-long-traces",
      "claim_type": "reproduced-calculation",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§4.3.4; Figure8"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "R2是最有解释力的失败：均分36.94→35.90。作者检查到42条长轨迹占135条的31.11%，却贡献84.2%的监督tokens。重复135条24次并不意味着所有任务均匀增加了监督，少数长轨迹会主导训练。R3扩展至558条首次使用轨迹并训练4 epochs，均分回到39.22；R4继续补约束检查与操作后状态读取，R5把203/332条新数据给BFCL，均分最终42.95。"
        }
      ]
    },
    {
      "id": "rsi-confounds",
      "claim_type": "interpretation",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "Figure8 caption; AppendixA.3"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "这段记录能展示Agent怎样发现数据配方失败并调整实验，但每轮同时改变数据、初始化、重复次数、epochs和steps。不能据此把R3回升单独归因于“更多独立样本”，或把轮数当作相同计算预算。R5的BFCL升幅9.07与集中补数据相符，仍不是独立因果消融。"
        }
      ]
    },
    {
      "id": "rsi-results-negative",
      "claim_type": "reproduced-calculation",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "Table2"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "按Table 2八项原数重算，R0均值33.1875，R5均值42.94875，增加9.76125个百分点，四舍五入对应33.19→42.95。最终7/8项高于R0，JobBench为22.66→20.53，下降2.13；GDPVal中间R2达到77.27，最终76.45。不能说每轮单调改善或所有任务提升。Automation的33.15和SkillsBench的14.91合计贡献八项总提升的61.54%，提升在任务间也不均匀。"
        }
      ]
    },
    {
      "id": "sampling-epoch-step",
      "claim_type": "author-claim",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "p34 unnumbered training table"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "p.34未编号表保留各项定义：R0/R1从base独立开始，分别66/69steps；R2–R5接续上轮，10/24/22/20steps。R2为1epoch，R3–R5为4epochs，新增数据24×与old 1×是采样乘数，不能与epochs混为一谈。原seed来源写作320条open-source BC、24条self-built BC、261条DSQA和两组627/113条sleep-filtered WS；原文没有在这里完整展开这些简称，不猜测对应数据集。"
        }
      ]
    },
    {
      "id": "release-status",
      "claim_type": "inspected-implementation",
      "source_ids": [
        "code"
      ],
      "source_locators": [
        {
          "source_id": "code",
          "locator": "package.json:1–24; tracked paths and README/docs/src/.pi/scripts search"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "公开实现是一套完整TypeScript/Electron框架，package版本0.1.3，包含CLI、桌面端、内置应用、运行时和测试，而不只是论文或静态demo。不过在本次固定revision的tracked文件和显式公开链接中，没有找到五轮SFT的实际数据、训练脚本、权重、trainer_state或对应checkpoint，也没有找到主表各benchmark的完整适配器和运行记录。框架可获取与实验可复现是两个不同结论；源码里的runtime checkpoint和mock测试不能填补模型训练发布物。"
        }
      ]
    },
    {
      "id": "gallery-review-denominators",
      "claim_type": "inspected-implementation",
      "source_ids": [
        "project",
        "review"
      ],
      "source_locators": [
        {
          "source_id": "project",
          "locator": "cases.json; review/protocol.json; checked 2026-11-09"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "官网画廊有24条case，作者公开review覆盖19案例、57份产物和92页Office，完整游戏通关为0次；Butterfly Life Cycle在walkthrough后加入，不属于那19例。它是作者发布的Agent reviewers检查，不是独立人类研究。本次只读公开记录，没有亲自运行这些产物，不能将24张好看示例写成24项正式成功任务。"
        }
      ]
    },
    {
      "id": "adaptive-evaluation-risk",
      "claim_type": "interpretation",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "§4.3; AppendixA.3"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "对RSI结果，还要保留评测适应风险：每轮持续根据同一组benchmark反馈选择训练内容，即使BM25剔除近重复任务，也不能排除对评测集合逐轮适应。缺少独立blind heldout、等预算人工数据配方、固定自动配方，以及更长SFT等对照，现有结果不足以建立自动化研究比这些路线更有效。固定教师生成演示具有行为模仿成分，但没有报告OPD所需的在线teacher分布或KL目标。"
        }
      ]
    },
    {
      "id": "future-work",
      "claim_type": "interpretation",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "Discussion grounded in §3–4 limitations"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "后续研究可以分别测试三个环节。共享资源用并发修改和过时候选检验是否保留人的新版本；持续执行用重启、消息重复、网络失败和等待条件消失检验恢复；自动化训练使用固定开发集选数据、独立隐藏测试集报告结果，并控制teacher调用、训练tokens和GPU预算。这些是本文提出的可证伪实验，不是作者已经做过的结果。"
        }
      ]
    },
    {
      "id": "conclusion",
      "claim_type": "interpretation",
      "source_ids": [
        "paper"
      ],
      "source_locators": [
        {
          "source_id": "paper",
          "locator": "Table1; Table2; AppendixA.3"
        }
      ],
      "evidence": [
        {
          "artifact": "article",
          "quote": "实验支持同一GLM模型下广泛的框架竞争力，并展示固定GLM组织Qwen五轮SFT后均分增加9.76个百分点。应同时记住R2下降、JobBench退步、任务增益集中，以及训练发布物和等预算对照缺失。对关注harness训练的人，这篇最有用的内容是任务状态、异步反馈、验收责任与数据配方失败的具体记录；它提供SFT研究过程案例，不能作为已经实现harness RL或agentic OPD的证据。"
        }
      ]
    }
  ],
  "source_visuals": [
    {
      "id": "figure-1",
      "kind": "figure",
      "label": "Figure 1",
      "source_id": "paper",
      "source_locator": "p.1, Figure 1",
      "status": "reused",
      "rationale": "持续工作与共同修改的完整总览。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 1，p.1。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。持续工作与共同修改的完整总览。"
        },
        {
          "artifact": "deck",
          "locator": "Figure 1"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-1.webp"
      ]
    },
    {
      "id": "figure-2",
      "kind": "figure",
      "label": "Figure 2",
      "source_id": "paper",
      "source_locator": "p.3, Figure 2",
      "status": "reused",
      "rationale": "项目工作和持续监督两类任务、程序监控与模型审阅。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 2，p.3。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。项目工作和持续监督两类任务、程序监控与模型审阅。"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-2.webp"
      ]
    },
    {
      "id": "figure-3",
      "kind": "figure",
      "label": "Figure 3",
      "source_id": "paper",
      "source_locator": "p.6, Figure 3",
      "status": "reused",
      "rationale": "Main、Coordinator与Subagents，以及共享应用资源和工作区。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 3，p.6。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。Main、Coordinator与Subagents，以及共享应用资源和工作区。"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-3.webp",
        "/lib/papers/coworkos/figure-3-runtime-detail.webp"
      ]
    },
    {
      "id": "figure-4",
      "kind": "figure",
      "label": "Figure 4",
      "source_id": "paper",
      "source_locator": "p.10, Figure 4",
      "status": "reused",
      "rationale": "任务目录、独立资源存储，以及房屋v1–v4的论文示意。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 4，p.10。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。任务目录、独立资源存储，以及房屋v1–v4的论文示意。"
        },
        {
          "artifact": "deck",
          "locator": "Figure 4"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-4.webp"
      ]
    },
    {
      "id": "figure-5",
      "kind": "figure",
      "label": "Figure 5",
      "source_id": "paper",
      "source_locator": "p.11, Figure 5",
      "status": "reused",
      "rationale": "Blender中的代码执行、文件更新与原生viewport；不是动画评测。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 5，p.11。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。Blender中的代码执行、文件更新与原生viewport；不是动画评测。"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-5.webp"
      ]
    },
    {
      "id": "figure-6",
      "kind": "figure",
      "label": "Figure 6",
      "source_id": "paper",
      "source_locator": "p.17, Figure 6",
      "status": "reused",
      "rationale": "六类产物完整对照；保留全部案例，不将视觉比较当任务成功率。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 6，p.17。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。六类产物完整对照；保留全部案例，不将视觉比较当任务成功率。"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-6.webp"
      ]
    },
    {
      "id": "figure-7",
      "kind": "figure",
      "label": "Figure 7",
      "source_id": "paper",
      "source_locator": "p.20, Figure 7",
      "status": "reused",
      "rationale": "八项R0–R5曲线，含GDPVal中间峰值和JobBench退步。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 7，p.20。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。八项R0–R5曲线，含GDPVal中间峰值和JobBench退步。"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-7.webp"
      ]
    },
    {
      "id": "figure-8",
      "kind": "figure",
      "label": "Figure 8",
      "source_id": "paper",
      "source_locator": "p.21, Figure 8",
      "status": "reused",
      "rationale": "完整均分与每轮训练配方，保留R2下降。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 8，p.21。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。完整均分与每轮训练配方，保留R2下降。"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-8.webp"
      ]
    },
    {
      "id": "figure-9",
      "kind": "figure",
      "label": "Figure 9",
      "source_id": "paper",
      "source_locator": "p.28, Figure 9",
      "status": "reused",
      "rationale": "Kitchen信息与Ridge Pack拒绝反馈；基线同样有实际交互。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 9，p.28。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。Kitchen信息与Ridge Pack拒绝反馈；基线同样有实际交互。"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-9.webp",
        "/lib/papers/coworkos/figure-9-ridge-detail.webp"
      ]
    },
    {
      "id": "figure-10",
      "kind": "figure",
      "label": "Figure 10",
      "source_id": "paper",
      "source_locator": "p.29, Figure 10",
      "status": "reused",
      "rationale": "Paper Glider与Fruit Slicer可见场景；各游戏数字不是统一成功率。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 10，p.29。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。Paper Glider与Fruit Slicer可见场景；各游戏数字不是统一成功率。"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-10.webp"
      ]
    },
    {
      "id": "figure-11",
      "kind": "figure",
      "label": "Figure 11",
      "source_id": "paper",
      "source_locator": "p.30, Figure 11",
      "status": "reused",
      "rationale": "Office内容页的指标布局和因果图；排版不保证内容数值正确。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 11，p.30。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。Office内容页的指标布局和因果图；排版不保证内容数值正确。"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-11.webp"
      ]
    },
    {
      "id": "figure-12",
      "kind": "figure",
      "label": "Figure 12",
      "source_id": "paper",
      "source_locator": "p.31, Figure 12",
      "status": "reused",
      "rationale": "交互可视化的细节检查；视觉可读性不等于算法验证。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 12，p.31。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。交互可视化的细节检查；视觉可读性不等于算法验证。"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-12.webp"
      ]
    },
    {
      "id": "figure-13",
      "kind": "figure",
      "label": "Figure 13",
      "source_id": "paper",
      "source_locator": "p.32, Figure 13",
      "status": "reused",
      "rationale": "发动机和机械关节的原生视图细节；几何导入不等于制造可行。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 13，p.32。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。发动机和机械关节的原生视图细节；几何导入不等于制造可行。"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-13.webp"
      ]
    },
    {
      "id": "figure-14",
      "kind": "figure",
      "label": "Figure 14",
      "source_id": "paper",
      "source_locator": "p.33, Figure 14",
      "status": "reused",
      "rationale": "宫殿和街景局部，完整保留CoworkOS的公交悬空反例。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Figure 14，p.33。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。宫殿和街景局部，完整保留CoworkOS的公交悬空反例。"
        },
        {
          "artifact": "deck",
          "locator": "Figure 14"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/figure-14.webp"
      ]
    },
    {
      "id": "table-1",
      "kind": "table",
      "label": "Table 1",
      "source_id": "paper",
      "source_locator": "p.15, Table 1",
      "status": "reused",
      "rationale": "十二项四框架完整结果；GDPVal Elo按原协议转换后进入领域均分。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Table 1，p.15。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。十二项四框架完整结果；GDPVal Elo按原协议转换后进入领域均分。"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/table-1.webp"
      ]
    },
    {
      "id": "table-2",
      "kind": "table",
      "label": "Table 2",
      "source_id": "paper",
      "source_locator": "p.33, Table 2",
      "status": "reused",
      "rationale": "R0–R5全部八项、均分和相对起点变化，含JobBench负增益。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文Table 2，p.33。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。R0–R5全部八项、均分和相对起点变化，含JobBench负增益。"
        },
        {
          "artifact": "deck",
          "locator": "Table 2"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/table-2.webp"
      ]
    },
    {
      "id": "training-table",
      "kind": "table",
      "label": "Unnumbered training table",
      "source_id": "paper",
      "source_locator": "p.34, Unnumbered training table",
      "status": "reused",
      "rationale": "p.34未编号训练表和seed说明；保留谱系、epochs、采样乘数与steps。",
      "placements": [
        {
          "artifact": "article",
          "locator": "手机可横向滑动或放大。 原论文p.34未编号训练表。CoworkOS Team，CC BY-NC-SA 4.0。p.34未编号训练表和seed说明；保留谱系、epochs、采样乘数与steps。"
        }
      ],
      "permission_status": "verified-allowed",
      "permission_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team, arXiv:2611.03178v1, CC BY-NC-SA 4.0",
      "asset_paths": [
        "/lib/papers/coworkos/training-table.webp"
      ]
    }
  ],
  "figures": [
    {
      "path": "/lib/papers/coworkos/figure-1.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-1",
      "source_id": "paper",
      "source_locator": "p.1, Figure 1",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 1，p.1。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。持续工作与共同修改的完整总览。",
      "used_in": [
        "article",
        "deck"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-2.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-2",
      "source_id": "paper",
      "source_locator": "p.3, Figure 2",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 2，p.3。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。项目工作和持续监督两类任务、程序监控与模型审阅。",
      "used_in": [
        "article"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-3.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-3",
      "source_id": "paper",
      "source_locator": "p.6, Figure 3",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 3，p.6。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。Main、Coordinator与Subagents，以及共享应用资源和工作区。",
      "used_in": [
        "article"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-4.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-4",
      "source_id": "paper",
      "source_locator": "p.10, Figure 4",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 4，p.10。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。任务目录、独立资源存储，以及房屋v1–v4的论文示意。",
      "used_in": [
        "article",
        "deck"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-5.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-5",
      "source_id": "paper",
      "source_locator": "p.11, Figure 5",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 5，p.11。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。Blender中的代码执行、文件更新与原生viewport；不是动画评测。",
      "used_in": [
        "article"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-6.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-6",
      "source_id": "paper",
      "source_locator": "p.17, Figure 6",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 6，p.17。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。六类产物完整对照；保留全部案例，不将视觉比较当任务成功率。",
      "used_in": [
        "article"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-7.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-7",
      "source_id": "paper",
      "source_locator": "p.20, Figure 7",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 7，p.20。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。八项R0–R5曲线，含GDPVal中间峰值和JobBench退步。",
      "used_in": [
        "article"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-8.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-8",
      "source_id": "paper",
      "source_locator": "p.21, Figure 8",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 8，p.21。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。完整均分与每轮训练配方，保留R2下降。",
      "used_in": [
        "article"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-9.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-9",
      "source_id": "paper",
      "source_locator": "p.28, Figure 9",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 9，p.28。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。Kitchen信息与Ridge Pack拒绝反馈；基线同样有实际交互。",
      "used_in": [
        "article"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-10.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-10",
      "source_id": "paper",
      "source_locator": "p.29, Figure 10",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 10，p.29。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。Paper Glider与Fruit Slicer可见场景；各游戏数字不是统一成功率。",
      "used_in": [
        "article"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-11.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-11",
      "source_id": "paper",
      "source_locator": "p.30, Figure 11",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 11，p.30。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。Office内容页的指标布局和因果图；排版不保证内容数值正确。",
      "used_in": [
        "article"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-12.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-12",
      "source_id": "paper",
      "source_locator": "p.31, Figure 12",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 12，p.31。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。交互可视化的细节检查；视觉可读性不等于算法验证。",
      "used_in": [
        "article"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-13.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-13",
      "source_id": "paper",
      "source_locator": "p.32, Figure 13",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 13，p.32。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。发动机和机械关节的原生视图细节；几何导入不等于制造可行。",
      "used_in": [
        "article"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-14.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-14",
      "source_id": "paper",
      "source_locator": "p.33, Figure 14",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Figure 14，p.33。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。宫殿和街景局部，完整保留CoworkOS的公交悬空反例。",
      "used_in": [
        "article",
        "deck"
      ]
    },
    {
      "path": "/lib/papers/coworkos/table-1.webp",
      "origin": "paper-original",
      "source_visual_id": "table-1",
      "source_id": "paper",
      "source_locator": "p.15, Table 1",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Table 1，p.15。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。十二项四框架完整结果；GDPVal Elo按原协议转换后进入领域均分。",
      "used_in": [
        "article"
      ]
    },
    {
      "path": "/lib/papers/coworkos/table-2.webp",
      "origin": "paper-original",
      "source_visual_id": "table-2",
      "source_id": "paper",
      "source_locator": "p.33, Table 2",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文Table 2，p.33。CoworkOS Team，arXiv:2611.03178v1，CC BY-NC-SA 4.0。R0–R5全部八项、均分和相对起点变化，含JobBench负增益。",
      "used_in": [
        "article",
        "deck"
      ]
    },
    {
      "path": "/lib/papers/coworkos/training-table.webp",
      "origin": "paper-original",
      "source_visual_id": "training-table",
      "source_id": "paper",
      "source_locator": "p.34, Unnumbered training table",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "手机可横向滑动或放大。 原论文p.34未编号训练表。CoworkOS Team，CC BY-NC-SA 4.0。p.34未编号训练表和seed说明；保留谱系、epochs、采样乘数与steps。",
      "used_in": [
        "article"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-3-runtime-detail.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-3",
      "source_id": "paper",
      "source_locator": "p.6, figure-3 labeled detail",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team，CC BY-NC-SA 4.0，原图局部，完整图保留在正文",
      "used_in": [
        "deck"
      ]
    },
    {
      "path": "/lib/papers/coworkos/figure-9-ridge-detail.webp",
      "origin": "paper-original",
      "source_visual_id": "figure-9",
      "source_id": "paper",
      "source_locator": "p.28, figure-9 labeled detail",
      "license_status": "verified-allowed",
      "license_basis": "arXiv v1 links CC BY-NC-SA 4.0; this attributed noncommercial paper explainer reproduces complete original figures/tables with source and license links, same license for cropped reproductions. Original English labels and numbers preserved.",
      "credit": "CoworkOS Team，CC BY-NC-SA 4.0，原图局部，完整图保留在正文",
      "used_in": [
        "deck"
      ]
    },
    {
      "path": "/lib/papers/coworkos/revision-flow.svg",
      "origin": "generated",
      "source_id": "paper",
      "source_locator": "§3.2.1 Eq7",
      "credit": "本文依据论文与官方代码绘制机制或数据关系，非执行日志",
      "used_in": [
        "article"
      ],
      "deck_usage": "SVG embedded inline"
    },
    {
      "path": "/lib/papers/coworkos/sft-flow.svg",
      "origin": "generated",
      "source_id": "paper",
      "source_locator": "§4.3; AppendixA.3",
      "credit": "本文依据论文与官方代码绘制机制或数据关系，非执行日志",
      "used_in": [
        "article"
      ],
      "deck_usage": "SVG embedded inline"
    },
    {
      "path": "/lib/papers/coworkos/rsi-trend.svg",
      "origin": "generated",
      "source_id": "paper",
      "source_locator": "Table2",
      "credit": "本文依据论文与官方代码绘制机制或数据关系，非执行日志",
      "used_in": [
        "deck"
      ],
      "deck_usage": "SVG embedded inline"
    }
  ],
  "qa": {
    "reading_layout": {
      "deck_first": true,
      "toc_h2_only": true,
      "surface": "neutral-light",
      "body_font": "serif",
      "body_font_px": 16,
      "line_height": 1.78,
      "figures_centered": true,
      "article_visuals_interleaved": true,
      "internal_scroll_for_tables_and_code": true,
      "page_horizontal_overflow": false
    },
    "viewports": [
      {
        "width": 1600,
        "height": 1000,
        "passed": true
      },
      {
        "width": 1512,
        "height": 982,
        "passed": true
      },
      {
        "width": 390,
        "height": 844,
        "passed": true
      }
    ],
    "tested_surfaces": {
      "article_body": true,
      "embedded_deck": true,
      "standalone_deck": true
    },
    "language_review": {
      "technical_terms_contextualized": true,
      "actors_disambiguated": true,
      "read_aloud_complete": true,
      "facts_rechecked_after_rewrite": true
    },
    "interactions": {
      "buttons": true,
      "keyboard": true,
      "hash_navigation": true,
      "fullscreen": true,
      "escape": true,
      "standalone_return_to_article": true,
      "immersive_mode_exit": true,
      "usable_state_after_exit": true
    },
    "inspected_slides": [
      "opening",
      "densest",
      "largest-table-or-code",
      "final"
    ],
    "evidence_paths": [
      "work/paper-2611-03178/qa-report.json",
      "work/paper-2611-03178/qa-*.png",
      "work/paper-2611-03178/qa-body-*.jpg"
    ],
    "scope": "Isolated headless Chromium; DOM fullscreen; no native desktop-window claim."
  }
}