宜博 / lemonboss

{
  "meta": {
    "task": "T-mtcng5rb / T-mtcnrsgk / T-mtco2k3h — deep-research Agent Hunt Phase 1",
    "agent": "innovator-hunt (Hunter)",
    "history_note": "本清单为同日两轮执行的合并结果。run1（2026-08-28 08:40 +08:00，另一实例）抓到 11 条但受当时网络出口限制，OpenAI/Gemini/SIFT 全部 blocked；run2（2026-08-28 08:15-08:37 UTC，本实例）网络窗口变好，把 run1 里 D 级未复检的项补成 A 级。两轮条目全部保留，未删任何一条。",
    "counts": {
      "成功缓存条数": 21,
      "其中有效信源": 19,
      "发现用辅助件": 2,
      "官方一手条数": 10,
      "验收线": "≥8 条缓存 / 官方一手 ≥3",
      "达标": true
    },
    "cross_check": "run1 经 cdn.jsdelivr.net 镜像抓的 4 个文件与 run2 经 raw.githubusercontent.com 直连抓的同一文件 sha256 逐一相同（gpt-researcher README、STORM README、Anthropic multi-agent 全文、MBFC methodology），镜像保真度已验证。",
    "cache_dir": "sources_raw/",
    "run1_网络": "raw.githubusercontent.com 不可达（走 jsdelivr 镜像）；openai.com 403；support.google.com / en.wikipedia.org / webliteracy.pressbooks.com 超时或 DNS 失败；docs.claude.com 部分页区域封锁",
    "run2_网络": "raw.githubusercontent.com / api.github.com 可达；anthropic.com / blog.google / arxiv.org / mediabiasfactcheck.com / hapgood.us / fullfact.org 可达；openai.com 仍 403；docs.perplexity.ai 超时；WebFetch 对部分域名报「Unable to verify if domain is safe to fetch」"
  },
  "sources": [
    {
      "id": "S1",
      "run": 1,
      "url": "https://www.anthropic.com/engineering/built-multi-agent-research-system",
      "title": "How we built our multi-agent research system（2025-06-13）",
      "publisher": "Anthropic Engineering Blog",
      "tier": "官方一手",
      "cache_file": "sources_raw/anthropic-multiagent.html",
      "bytes": 205221,
      "sha256": "b61b8b20023e54ec94e4ca78bb6202503c14de93b04466594d17954ca61ca956",
      "抓取方式": "run1: curl -sS -L --max-time 40 --retry 1（UA Mozilla/5.0）；run2 复抓同一 URL 得字节级相同内容（sha256 一致）",
      "note": "本轮信息密度最高的官方一手来源：orchestrator-worker、effort scaling rules、先宽后窄、LLM-as-judge 五维 rubric、CitationAgent、反模式清单、token 4×/15×"
    },
    {
      "id": "S2",
      "run": 1,
      "url": "https://www.anthropic.com/engineering/writing-tools-for-agents",
      "title": "Writing effective tools for AI agents—using AI agents",
      "publisher": "Anthropic Engineering Blog",
      "tier": "官方一手",
      "cache_file": "sources_raw/anthropic-engineering-writing-tools.html",
      "bytes": 233652,
      "sha256": "fc7cf0e78287aa867373dec83b16c73ce13f9c62826b5aff8e41c6750729b9e5",
      "抓取方式": "run1: curl -sS -L --max-time 40 --retry 1",
      "note": "工具描述质量决定 agent 成败；自测试 agent 改写工具描述后任务完成时间降 40%"
    },
    {
      "id": "S3",
      "run": 1,
      "url": "https://code.claude.com/docs/en/skills",
      "title": "Extend Claude with skills（Claude Code 官方文档）",
      "publisher": "Anthropic 官方文档",
      "tier": "官方一手",
      "cache_file": "sources_raw/claude-code-skills-docs.html",
      "bytes": 1058896,
      "sha256": "57634bb5bc01113620231191fd0fea204647276972048a7a4d352165bffb938f",
      "抓取方式": "run1: curl -sS -L --max-time 40 --retry 1",
      "note": "含官方 deep-research skill 字面模板（context: fork + agent: Explore）与 description 1536 字符上限等工程硬约束"
    },
    {
      "id": "S4",
      "run": 1,
      "url": "https://github.com/anthropics/skills/blob/main/README.md",
      "title": "anthropics/skills README",
      "publisher": "Anthropic 官方仓库",
      "tier": "官方一手",
      "cache_file": "sources_raw/anthropic-skills-repo-readme.md",
      "bytes": 5552,
      "sha256": "2fb9c4cc026366ea6e21c60cb23c908a4d27c62071819cc06beffaac102d0694",
      "抓取方式": "run1: jsdelivr 镜像；run2 直连 raw.githubusercontent.com 同一文件 sha256 一致",
      "note": "SKILL.md 结构 / Apache 2.0 范围 / 技能打包模式"
    },
    {
      "id": "S5",
      "run": 1,
      "url": "https://github.com/anthropics/claude-code/blob/main/README.md",
      "title": "anthropics/claude-code README",
      "publisher": "Anthropic 官方仓库",
      "tier": "官方一手",
      "cache_file": "sources_raw/anthropic-claude-code-repo-readme.md",
      "bytes": 2873,
      "sha256": "aa0d8b80ac083e1591be0789273f447058cc9187b4331634a5614155da27778d",
      "抓取方式": "run1: jsdelivr 镜像",
      "note": "运行时定位：deep-research skill 跑在 Claude Code agent loop 上，非独立服务"
    },
    {
      "id": "S6",
      "run": 1,
      "url": "https://github.com/anthropics/anthropic-cookbook/blob/main/README.md",
      "title": "Claude Cookbooks README",
      "publisher": "Anthropic 官方仓库",
      "tier": "官方一手",
      "cache_file": "sources_raw/anthropic-cookbook-readme.md",
      "bytes": 6083,
      "sha256": "5d5af555aff7b94e67b38478e20463c09aaa1c6d7202b75d91ee1eb49a262a41",
      "抓取方式": "run1: jsdelivr 镜像",
      "note": "官方示例仓库索引（辅助）"
    },
    {
      "id": "S7",
      "run": 1,
      "url": "https://github.com/assafelovic/gpt-researcher/blob/master/README.md",
      "title": "GPT Researcher README",
      "publisher": "开源项目（Assaf Elovic）",
      "tier": "开源仓库",
      "cache_file": "sources_raw/gpt-researcher-readme.md",
      "bytes": 16874,
      "sha256": "9f8cb53ece3fd00740c3a5cf759406547fe49c8b5d2813201ded8bd3db05e1d1",
      "抓取方式": "run1: jsdelivr 镜像；run2 直连 raw.githubusercontent.com 同一文件 sha256 一致",
      "note": "planner/execution/publisher 三段式 + 五步流程 + 聚合 20+ 源、报告 >2000 词、多源取众数降错率"
    },
    {
      "id": "S8",
      "run": 1,
      "url": "https://github.com/stanford-oval/storm/blob/main/README.md",
      "title": "STORM README（Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking）",
      "publisher": "Stanford OVAL（开源）",
      "tier": "开源仓库",
      "cache_file": "sources_raw/storm-readme.md",
      "bytes": 20091,
      "sha256": "711016d02ae72a43ae96198ede47ee0b36ab10029b323a76eb2d3cd87413e2a2",
      "抓取方式": "run1: jsdelivr 镜像；run2 直连 raw.githubusercontent.com 同一文件 sha256 一致",
      "note": "pre-writing/writing 两阶段 + Perspective-Guided Question Asking + Simulated Conversation 逐字命中"
    },
    {
      "id": "S9",
      "run": 1,
      "url": "https://arxiv.org/abs/2402.14207",
      "title": "STORM 论文摘要页（NAACL 2024）",
      "publisher": "arXiv",
      "tier": "学术一手",
      "cache_file": "sources_raw/storm-paper-arxiv.html",
      "bytes": 42499,
      "sha256": "23df7d11895bba2a0ee747c23ddd86e7bb6fe73791aea1e64abbb148ee82d61c",
      "抓取方式": "run1: curl 直连；run2 复抓同一 URL sha256 一致",
      "note": "与 S10/S17 同一论文"
    },
    {
      "id": "S10",
      "run": 1,
      "url": "https://arxiv.org/pdf/2402.14207",
      "title": "STORM 论文全文 PDF（27pp）",
      "publisher": "arXiv",
      "tier": "学术一手",
      "cache_file": "sources_raw/storm-paper-arxiv-pdf.pdf",
      "bytes": 1320410,
      "sha256": "dae3d662c85488ce41f3fbae8f66224431acb8ab6a5a5fb3ae3bb5c84b0d0f0b",
      "抓取方式": "run1: curl 直连",
      "note": "source bias transfer / over-association of unrelated facts 两个学术命名的编造类缺陷出自此文"
    },
    {
      "id": "S11",
      "run": 1,
      "url": "https://mediabiasfactcheck.com/methodology/",
      "title": "Our Methodology（Media Bias/Fact Check，2025-01-01 生效）",
      "publisher": "Media Bias/Fact Check",
      "tier": "第三方方法论",
      "cache_file": "sources_raw/mbfc-methodology.html",
      "bytes": 223101,
      "sha256": "343b9d26c1ada94425c2cdaef63331e3b3f50ee264bd79fa9f5fee9f276ad47a",
      "抓取方式": "run1: curl 直连；run2 复抓同一 URL sha256 一致",
      "note": "唯一可复算权重的信源可信度框架：偏见四维 35/35/15/15（±10 分 7 档）、事实性四维 40/25/25/10（6 档）、最低样本 10 标题+5 全文"
    },
    {
      "id": "S12",
      "run": 2,
      "url": "https://raw.githubusercontent.com/openai/openai-cookbook/main/examples/data/openai_blog_pdfs/Introducing%20deep%20research%20_%20OpenAI.pdf",
      "title": "Introducing deep research（OpenAI 官方博客全文 PDF，2025-02-02）",
      "publisher": "OpenAI",
      "tier": "官方一手",
      "cache_file": "sources_raw/openai__introducing-deep-research-blog.pdf",
      "bytes": 1086209,
      "sha256": "ad778baeea27fbd6dd2e69d354a8d437d4b9985c70836031e0712ba1d5b0ef67",
      "抓取方式": "run2: curl --retry 3 --retry-all-errors --max-time 60（UA Mozilla/5.0），HTTP 200；路径由 S20 文件树枚举定位",
      "note": "openai.com 原站 403（见 blocked），改抓 OpenAI 官方 cookbook 仓库自带镜像，内容等价；run1 的最大缺口由此补上"
    },
    {
      "id": "S13",
      "run": 2,
      "url": "https://raw.githubusercontent.com/openai/openai-cookbook/main/examples/data/openai_blog_pdfs/Deep%20research%20System%20Card%20_%20OpenAI.pdf",
      "title": "Deep research System Card（OpenAI，2025-02-25）",
      "publisher": "OpenAI",
      "tier": "官方一手",
      "cache_file": "sources_raw/openai__deep-research-system-card.pdf",
      "bytes": 117014,
      "sha256": "49b6b933cbe076cac3bfe52b2186acb9e3e76fc4a61d18f824628c3767755230",
      "抓取方式": "run2: curl --retry 3 --retry-all-errors --max-time 60，HTTP 200",
      "note": "multi-step research on the internet / 优化 web browsing / Preparedness Scorecard（post-mitigation ≤ medium 才可部署）"
    },
    {
      "id": "S14",
      "run": 2,
      "url": "https://raw.githubusercontent.com/openai/openai-cookbook/main/examples/deep_research_api/how_to_build_a_deep_research_mcp_server/README.md",
      "title": "MCP for Deep Research（OpenAI Cookbook）",
      "publisher": "OpenAI",
      "tier": "官方一手",
      "cache_file": "sources_raw/openai__deep-research-mcp-server-README.md",
      "bytes": 3788,
      "sha256": "d1652c403ede0bfaad0bdf8a55313ac14b202c58bc52157f66607ae0c45b3821",
      "抓取方式": "run2: curl --retry 3 --retry-all-errors --max-time 60，HTTP 200",
      "note": "官方明文：Deep Research agent 依赖 Search + Fetch 两工具；README 内嵌官方 system_message（数据密集、优先可靠一手源、inline citations、已取过的文件不重复 fetch）"
    },
    {
      "id": "S15",
      "run": 2,
      "url": "https://blog.google/products/gemini/google-gemini-deep-research/",
      "title": "Gemini: Deep Research（Google 官方博客，2024-12）",
      "publisher": "Google",
      "tier": "官方一手",
      "cache_file": "sources_raw/google__gemini-deep-research.md",
      "bytes": 372154,
      "sha256": "b3ba119859bc1d2699f55ede69016bdb6efc7c15e4414c8cf10582ad160cc03a",
      "抓取方式": "run2: curl --retry 3 --retry-all-errors --max-time 60，HTTP 200",
      "note": "research plan 先给用户 revise 或 approve（human-in-the-loop）+ comprehensive report with key findings and links to original sources"
    },
    {
      "id": "S16",
      "run": 2,
      "url": "https://hapgood.us/2019/06/19/sift-the-four-moves/",
      "title": "SIFT (The Four Moves) — Mike Caulfield（SIFT 原始出处）",
      "publisher": "hapgood.us",
      "tier": "权威实践方法论",
      "cache_file": "sources_raw/hapgood__SIFT-four-moves.md",
      "bytes": 1150636,
      "sha256": "c0ceb277eea5709942ec0117ba2c76efa299d4526b677aa8820e62a2951ea6be",
      "抓取方式": "run2: curl --retry 3 --retry-all-errors --max-time 60，HTTP 200",
      "note": "run1 抓的是 2019-10-23 错误日期路径（404 软 404）；正确路径 2019-06-19，四动词原文逐字命中，run1 的 D 级综述由此升级 A 级"
    },
    {
      "id": "S17",
      "run": 2,
      "url": "https://arxiv.org/html/2402.14207v2",
      "title": "STORM 论文全文（HTML 版）",
      "publisher": "arXiv",
      "tier": "学术一手",
      "cache_file": "sources_raw/storm__full-paper.html",
      "bytes": 498578,
      "sha256": "d35a59306dad5ee4273ecc5307c00ff864dcf23c1312557e60e2ebe86bce4076",
      "抓取方式": "run2: curl --retry 3 --retry-all-errors --max-time 60，HTTP 200",
      "note": "citation recall/precision 定义（引 Gao et al. 2023）与数值 84.83/85.18、1-5 分 Broad Coverage rubric 逐字命中"
    },
    {
      "id": "S18",
      "run": 2,
      "url": "https://raw.githubusercontent.com/langchain-ai/open_deep_research/main/README.md",
      "title": "Open Deep Research README（LangChain）",
      "publisher": "langchain-ai/open_deep_research（开源）",
      "tier": "开源仓库",
      "cache_file": "sources_raw/open_deep_research__README.md",
      "bytes": 10265,
      "sha256": "0fc75629079c025b0200e6032a038edf9c91c332d4ce925cbc6165910008267f",
      "抓取方式": "run2: curl --retry 3 --retry-all-errors --max-time 60，HTTP 200",
      "note": "四段流水线 Summarization / Research / Compression / Final Report 逐字命中；Deep Research Bench + RACE 评分；实测成本 $45.98-$187.09/百题"
    },
    {
      "id": "S19",
      "run": 2,
      "url": "https://fullfact.org/toolkit/",
      "title": "Misinformation Toolkit（Full Fact，英国独立事实核查慈善机构）",
      "publisher": "Full Fact",
      "tier": "第三方方法论",
      "cache_file": "sources_raw/fullfact__how-to-fact-check.md",
      "bytes": 57179,
      "sha256": "1940def960a4396c4011abb65a894a29ffbe4740a3c5e111475d0afca82c1cd7",
      "抓取方式": "run2: curl --retry 3 --retry-all-errors --max-time 60，HTTP 200",
      "note": "限缩使用：该页是索引导航页（纯文本仅 5.2KB），列分主题核查指南（图片/视频/民调/标题），不含逐步方法论文本，不得冒充完整方法论"
    },
    {
      "id": "S20",
      "run": 2,
      "url": "https://api.github.com/repos/openai/openai-cookbook/git/trees/main?recursive=1",
      "title": "openai/openai-cookbook 完整文件树（发现用辅助件）",
      "publisher": "GitHub API",
      "tier": "辅助件",
      "cache_file": "sources_raw/openai__cookbook-tree.json",
      "bytes": 1147134,
      "sha256": "0cb951952df6568574f341a85cfbe6dcadb34a75711f575f8a206f22d9f027fd",
      "抓取方式": "run2: curl --retry 3 --retry-all-errors --max-time 60，HTTP 200",
      "note": "用于定位 deep research 相关官方文件精确路径，产出 S12/S13/S14"
    },
    {
      "id": "S21",
      "run": 2,
      "url": "https://api.github.com/repos/anthropics/skills/git/trees/main?recursive=1",
      "title": "anthropics/skills 完整文件树（509 项，发现用辅助件）",
      "publisher": "GitHub API",
      "tier": "辅助件",
      "cache_file": "sources_raw/anthropic__skills-repo-tree.json",
      "bytes": 159191,
      "sha256": "1e23ecbc6b0d3d4ee2c08cc8066d3f3de5ecff1220dd3551dbf5a3a1726b943a",
      "抓取方式": "run2: curl --retry 3 --retry-all-errors --max-time 60，HTTP 200",
      "note": "用于穷举判定「官方是否已发布 web-research/deep-research 技能」——结论见 negative_findings 第 1 条"
    }
  ],
  "negative_findings": [
    {
      "id": "NF1",
      "claim": "anthropics/skills 仓库（509 个路径全量枚举）不存在 web-research / deep-research / fact-check 类技能",
      "evidence": "S21 文件树全量 grep research|search|fact 零命中",
      "implication": "「官方 deep-research SKILL.md 全文」不存在于公开仓库；S3（code.claude.com 文档页）里的 deep-research 模板是文档示例，不是仓库里的现役技能文件。两者不矛盾，引用时要分清出处"
    },
    {
      "id": "NF2",
      "claim": "OpenAI cookbook 没有 articles/deep_research.md 这类独立方法论文章",
      "evidence": "run2 猜测路径 curl 返回 404: Not Found（14 字节）；S20 全树枚举确认 deep research 内容只落在博客 PDF、系统卡 PDF 与 API 示例 notebook",
      "implication": "OpenAI 一手方法论以 S12/S13/S14 为准"
    },
    {
      "id": "NF3",
      "claim": "Gemini Deep Research 官方博客 HTML 中没有「browses hundreds of websites」这类数值表述",
      "evidence": "S15 缓存全文 grep hundreds 零命中；同类表述只出现在 OpenAI S12（synthesize hundreds of online sources）",
      "implication": "把「浏览数百个网站」挂在 Google 头上属误引，Google 官方明文只有 multi-step research plan + comprehensive report with key findings and links"
    },
    {
      "id": "NF4",
      "claim": "Full Fact toolkit 页不含可操作的逐步事实核查方法论文本",
      "evidence": "S19 提取纯文本仅 5224 字符，主体是分主题指南链接列表",
      "implication": "交叉验证章节只能引 Full Fact 作为核查实践存在的证据，不能逐字引其步骤"
    },
    {
      "id": "NF5",
      "claim": "没有任何官方信源定义「深度调研报告的标准结构」",
      "evidence": "S1/S8/S12/S14/S15/S18 逐字扫描，各系统只描述自己的输出形态",
      "implication": "报告结构范式属跨系统收敛实践 + 构造性设计，写进 SKILL 不得冒充官方判据"
    },
    {
      "id": "NF6",
      "claim": "多源冲突仲裁无可引官方条款（run1 已发现，run2 复核维持）",
      "evidence": "新增 8 条缓存（含 OpenAI 系统卡、Gemini 博客）逐字扫描后仍未找到任何一家系统化表述冲突仲裁规则",
      "implication": "仍是唯一需要上级拍板的裁决点之一"
    },
    {
      "id": "NF7",
      "claim": "CRAAP Test 的权威一手页面（CSU Chico Meriam Library）两轮都未能落盘全文",
      "evidence": "run1 两条路径 404；run2 换 /help/craap-test 仍 404（58KB 错误页），WebFetch 域名校验失败；五准则全文仅经 WebSearch 摘要获得（Currency/Relevance/Authority/Accuracy/Purpose + 1-4 打分 45-50 excellent 量表），属 C 级证据",
      "implication": "CRAAP 只能以「框架名 + 五准则」引用，禁引逐字题项"
    },
    {
      "id": "NF8",
      "claim": "LangChain 官方博客 URL 不可达",
      "evidence": "run2 curl blog.langchain.dev/langchain-deep-research/ 返回 HTTP 404（62KB 错误页，未入缓存）",
      "implication": "LangChain 一手材料以 S18 仓库 README 为准"
    }
  ],
  "blocked_sources": [
    {
      "url": "https://openai.com/index/introducing-deep-research/",
      "两轮症状": "run1/run2 均 curl HTTP 403（bot 拦截）；run2 降级 WebFetch 一次，返回 Unable to verify if domain openai.com is safe to fetch",
      "处置": "记入 blocked；run2 改抓 OpenAI 官方 cookbook 仓库镜像（S12/S13），内容等价，缺口已闭合"
    },
    {
      "url": "https://docs.perplexity.ai/getting-started/overview（及 /guides/getting-started）",
      "两轮症状": "run1 www.perplexity.ai TCP connect timeout ×2；run2 docs.perplexity.ai 连续 2 次 rc=28 超时（--retry 3 --retry-all-errors --max-time 60 用尽，文件未落盘），WebFetch 域名校验失败",
      "处置": "记入 blocked，标「无法复检」；Perplexity Pro Search 方法论在本次调研持续缺位"
    },
    {
      "url": "https://support.google.com/gemini/answer/15719131",
      "两轮症状": "run1 timeout 45s/40s ×3；run2 未再试（官方主信源已由 S15 补齐，帮助页仅补充价值）",
      "处置": "记入 blocked"
    },
    {
      "url": "https://docs.claude.com/en/docs/agents-and-tools/tool-use/web-search-tool 与 /en/docs/about-claude/research",
      "两轮症状": "HTTP 200 但正文为 App unavailable in region 区域封锁页（run1 两条 URL 返回字节级相同内容；run2 复测第一条同样如此），无 API 文档内容",
      "处置": "记入 blocked，标「无法复检」；run2 已删除 run2 抓到的占位副本以免污染缓存，run1 的原件保留在 sources_raw/ 供复核"
    },
    {
      "url": "https://en.wikipedia.org/wiki/SIFT_(method)",
      "两轮症状": "run1 TCP connect timeout ×2；run2 未再试（SIFT 原始出处已由 S16 补齐）",
      "处置": "记入 blocked"
    },
    {
      "url": "https://webliteracy.pressbooks.com/",
      "两轮症状": "run1 DNS 解析失败；run2 未再试（S16 已覆盖）",
      "处置": "记入 blocked"
    },
    {
      "url": "https://library.csuchico.edu/... 与 /help/craap-test（CRAAP 官方页）",
      "两轮症状": "run1 两处 404；run2 换路径仍 404，WebFetch 域名校验失败",
      "处置": "记入 blocked，标「无法复检」；内容经 WebSearch 摘要部分恢复（见 NF7），证据等级 C"
    },
    {
      "url": "https://raw.githubusercontent.com/*（run1 当时）",
      "两轮症状": "run1 直连 timeout（0 字节），改走 cdn.jsdelivr.net/gh/<org>/<repo>@<branch>/<path> 镜像全部 200；run2 直连恢复正常",
      "处置": "网络窗口差异记录在案；后续轮次应先试直连、失败再换镜像（镜像保真度已由 4 个文件 sha256 双向比对验证）"
    }
  ]
}