knowhow_v1.md 32.5 KB

Hunt Phase 1 调研报告 — deep-research Agent(深度调研方法论)

  • 任务: T-mtcng5rb(催办单 T-mtcnrsgk / T-mtco2k3h),org-innovator 2026-08-28 07:47 派发,master 07:56 催办
  • 执行: innovator-hunt (Hunter),2026-08-28 同日两轮合并交付:run1(08:40 +08:00,受当时网络出口限制,11 条缓存,OpenAI/Gemini/SIFT 不可达只能给 D 级);run2(08:15–08:37 UTC,网络窗口变好,补抓 8 条把 run1 的 D 级项升级成 A 级)。两轮条目全部保留,本文档 §1–§8 为 run1 主体,§9 为 run2 增量复核与 D→A 升级,冲突处以 §9 为准
  • 信源清单: sources_v1.json(合并后 21 条成功缓存,其中有效信源 19 条、官方一手 10 条;每条含 sha256 / 字节数 / 抓取方式 / run 归属)
  • 原文缓存: sources_raw/(SOP ②:首抓即落盘,后续轮次只读缓存禁复抓;run1 经 jsdelivr 镜像抓的 4 个文件与 run2 直连 raw.githubusercontent.com 抓的同一文件 sha256 逐一相同,镜像保真度已验证)
  • 配置复用: crawl_config.json(run1 实测抓取策略;run2 补充「先试直连、失败换镜像」的窗口差异,见 §9.5)

证据分级(本报告所有结论按此标注,禁止跨级引用)

级别 含义
A 官方一手逐字引文,出处精确到文件与 URL(S1–S6)
B 可操作的第三方公开方法论,有逐字引文(S7–S11)
C 构造性设计(我们自己提出的可训练规范),不是任何官方条款;另含仅经搜索摘要获得、原文未落盘的条目(如 CRAAP),不得当官方条款引用
D 常识性综述,原文未复检(信源不可达),不得当官方条款引用

run1 最大的诚实缺口是 OpenAI / Gemini / Perplexity / SIFT / CRAAP 五项全部 D 级; run2 已把其中 OpenAI(博客 PDF + 系统卡 + MCP README)、Gemini(官方博客)、SIFT(原始出处) 三项升级为 A 级(§9),仍留 D/C 级的是 Perplexity(持续不可达)与 CRAAP(官方页两轮 404)。


§0 结论速览

# 问题 结论 级别
1 deep-research 的主流架构 三种同型:多智能体 orchestrator-worker(Anthropic)、planner/execution+publisher(GPT Researcher)、两阶段 pre-writing/writing(STORM) A/B
2 并行策略 lead 并行起 3–5 个 subagent、每个 subagent 并行发 3+ 工具调用;复杂查询时延降 90% A
3 工作量分配 官方给出的 scaling rules:单点事实=1 agent/3–10 次调用;对比=2–4 subagent/各 10–15 次;复杂研究=10+ subagent A
4 查询策略 先宽后窄:先短查询探面,再逐步收窄;过早写长查询是官方点名失败模式 A
5 拆解质量判据 每个任务必须带 objective + output format + 工具/信源指引 + 任务边界,否则 subagent 重复劳动 A
6 引用与事实核查 报告生成后由独立 CitationAgent 把每条 claim 对回文档;评四个维度:事实/引用/完整/信源质量 A
7 信源评级可操作标准 MBFC 2025 版:偏见 4 维加权(35/35/15/15,±10 分 7 档);事实性 4 维加权(失败 fact check 40%/sourcing 25%/transparency 25%/one-sidedness 10%,6 档);最低样本 10 标题+5 全文 B
8 SIFT / CRAAP SIFT 原始出处 run2 已抓到原文(四动词逐字命中,A 级 §9.4);CRAAP 官方页两轮 404,仍只到 C 级(五准则经搜索摘要,禁引逐字题项) A / C
9 已知反模式 滥开 subagent(简单题开 50 个)、无限搜不存在的源、SEO 内容农场偏好、串行搜索、模糊派单、上下文溢出丢计划 A
10 与本组织的接合点 Claude Code 官方文档里就有 deep-research skill 同型模板(context: fork + agent: Explore),可直接作 SKILL 骨架母本 A

§1 主流深度调研系统的公开方法论

1.1 Anthropic 多智能体 Research 系统(S1,官方一手,最核心)

架构 = orchestrator-worker(lead agent + 并行 subagent)+ 独立 CitationAgent,流程:

"When a user submits a query, the system creates a LeadResearcher agent that enters an iterative research process. The LeadResearcher begins by thinking through the approach and saving its plan to Memory to persist the context… It then creates specialized Subagents… Each Subagent independently performs web searches, evaluates tool results using interleaved thinking, and returns findings to the LeadResearcher. The LeadResearcher synthesizes these results and decides whether more research is needed—if so, it can create additional subagents or refine its strategy. Once sufficient information is gathered, the system exits the research loop and passes all findings to a CitationAgent, which processes the documents and research report to identify specific locations for citations. This ensures all claims are properly attributed to their sources." — S1

要点(全部 A 级,逐字):

  • 线性单发流水线不适用:"Research demands the flexibility to pivot or explore tangential connections as the investigation unfolds… A linear, one-shot pipeline cannot handle these tasks."
  • 多智能体的本质是压缩与并行:"The essence of search is compression: distilling insights from a vast corpus. Subagents facilitate compression by operating in parallel with their own context windows… Each subagent also provides separation of concerns."
  • 收益量化:"a multi-agent system with Claude Opus 4 as the lead agent and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval."
  • token 决定论:"token usage by itself explains 80% of the variance"(BrowseComp);代价是 "multi-agent systems use about 15× more tokens than chats"。
  • 适用边界(训练时要写进 SKILL 的取舍条款):"multi-agent systems excel at valuable tasks that involve heavy parallelization, information that exceeds single context windows, and interfacing with numerous complex tools."

1.2 GPT Researcher(S7,开源一手)

  • 架构 = planner + execution + publisher:"The core idea is to utilize 'planner' and 'execution' agents. The planner generates research questions, while the execution agents gather relevant information. The publisher then aggregates all findings into a comprehensive report."
  • 五步流水线:"Create a task-specific agent based on a research query. → Generate questions that collectively form an objective opinion on the task. → Use a crawler agent for gathering information for each question. → Summarize and source-track each resource. → Filter and aggregate summaries into a final research report."
  • 深度模式 = 递归树:"Deep Research - an advanced recursive research workflow that explores topics with agentic depth and breadth… Tree-like exploration with configurable depth and breadth."
  • 对偏差的立场(训练反编造时可引):"We assume that the more sites we scrape the less chances of incorrect data. By scraping multiple sites per research, and choosing the most frequent information, the chances that they are all wrong is extremely low."(多源取众数降错率 —— 这是最朴素的交叉验证实现,B 级)

1.3 STORM / Co-STORM(S8 README + S9/S10 论文,学术一手)

  • 两阶段:Pre-writing(联网收集参考文献 + 生成 outline)→ Writing(据 outline+references 生成带引用全文)。"STORM breaks down generating long articles with citations into two steps."(S8)
  • 提问自动化 = 核心难题的两个解法:Perspective-Guided Question Asking(从同类主题既有文章挖多视角)+ Simulated Conversation(模拟 Wikipedia 写作者对基于可信网源的 topic expert 提问、追问)。"Directly prompting the language model to ask questions does not work well."(S8)
  • 四模块流水线(可训练骨架):Knowledge Curation → Outline Generation → Article Generation → Article Polishing(S8)
  • 效果与遗留问题:与 outline-driven RAG 基线比,"more of STORM's articles are deemed to be organized (by a 25% absolute increase) and broad in coverage (by 10%)";专家反馈点名两个新问题 "source bias transfer and over-association of unrelated facts"(信源偏差迁移 / 无关事实过度关联)——这两条就是「编造/污染」的学术命名,写 SKILL 反模式时直接引用(S9)。

1.4 OpenAI / Gemini / Perplexity(run1 时全 D 级;run2 已把 OpenAI、Gemini 升 A,见 §9)

run1 时三家官网全部不可达(§8),以下当时仅为常识综述、无任何逐字引文支撑:

  • OpenAI Deep Research(o3 驱动):单 agent 长程自主浏览 + 中途产出结构化引用报告,通常 5–30 分钟。→ run2 已抓到官方博客全文 PDF 与系统卡(S12/S13/S14),§1.4 的 OpenAI 条目整体作废,以 §9.1–§9.3 的 A 级引文为准。
  • Gemini Deep Research:先给用户一份 research plan 供确认,再执行多轮浏览。→ run2 已抓到 Google 官方博客(S15),「research plan 供 revise/approve」已是 A 级,见 §9.3。
  • Perplexity Pro Search:快速检索聚合,深度低于前两者。→ 仍为 D 级:run1 www.perplexity.ai TCP 超时 ×2,run2 docs.perplexity.ai 连续两次 rc=28 超时 + WebFetch 域名校验失败,标「无法复检」。 → Digester 对 Perplexity 的任何机制表述仍必须先拿原文,否则一律标 D。

§2 检索规划方法

2.1 需求澄清(A + C)

  • A(S1): lead agent 先分析 query、制定策略、把计划写入外部 Memory("saving its plan to Memory to persist the context, since if the context window exceeds 200,000 tokens it will be truncated and it is important to retain the plan")——长任务防丢计划是官方明确条款。
  • A(S15,run2 补强):Google 官方明文 "After you enter your question, it creates a multi-step research plan for you to either revise or approve."——「检索计划先给用户确认」从 D 级升级为 A 级,可直接作 SKILL 的 human-in-the-loop 判据。
  • C(构造性设计,供 SKILL 采纳):需求不清时先输出「检索计划」再动手——目标 / 已知 / 未知 / 子问题清单 / 信源优先级 / 预算(步数与时间)。依据:S1 的 effort scaling + S15 的 plan-confirm 机制 + STORM 的 pre-writing 阶段思想;「计划的具体字段集」本身是我们自定规范,不是官方条款。

2.2 子问题拆解(A)

  • 派单必须带四要素:"Each subagent needs an objective, an output format, guidance on the tools and sources to use, and clear task boundaries. Without detailed task descriptions, agents duplicate work, leave gaps, or fail to find necessary information."(S1)
  • 官方实测的翻车样例:"one subagent explored the 2021 automotive chip crisis while 2 others duplicated work investigating current 2025 supply chains, without an effective division of labor."(S1)
  • 多视角拆解(B):STORM 的 perspective-guided asking —— 先挖「谁会关心这个问题」的视角集合,再按视角提问,天然避免子问题同质化(S8)。

2.3 查询改写(A)

  • 先宽后窄:"start with short, broad queries, evaluate what's available, then progressively narrow focus",并明确点名反模式 "Agents often default to overly long, specific queries that return few results."(S1)
  • 追问式细化:STORM 的 simulated conversation 就是为 "update its understanding of the topic and ask follow-up questions"(S8)。
  • 工具选择启发(A,S1):"examine all available tools first, match tool usage to user intent, search the web for broad external exploration, or prefer specialized tools over generic ones."

2.4 多源并行检索策略(A)

  • 两种并行:"the lead agent spins up 3-5 subagents in parallel rather than serially; (2) the subagents use 3+ tools in parallel. These changes cut research time by up to 90% for complex queries"(S1)
  • 混合检索源(B,S7):GPT Researcher 支持 RETRIEVER=tavily,mcp(web + 私有库混合);STORM 实现了 You/Bing/Serper/Brave/SearXNG/DuckDuckGo/Tavily/Google/Azure 等多种 retriever 与 VectorRM 私有语料 grounding(S8)。→ 可训练成「公网检索 + 本地知识库」双通道。
  • 覆盖/收敛判据(A,S1):失败模式包括 "agents continuing when they already had sufficient results"(够用不停)与 "scouring the web endlessly for nonexistent sources"(找不到硬找)——SKILL 必须给出停机判据。

§3 信源可信度评级框架

3.1 MBFC 2025-01-01 生效方法论(S11,B 级,最可操作)

抽样底线:"our methodology requires reviewing a minimum of 10 headlines and 5 full news stories from each source."(可扩展至 About 页、更正记录、所有权文件、第三方报道、存档内容)

偏见评级 = 四维加权,±10 分制: | 维度 | 权重 | |---|---| | Economic System | 35% | | Social Progressive Liberalism vs. Traditional Social Conservatism | 35% | | Straight News Reporting Balance | 15% | | Editorial Bias | 15% |

分档:Extreme Left(−10~−8.0)/ Left(−7.9~−5.0)/ Left-Center(−4.9~−2.0)/ Least Biased(−1.9~+1.9) / Right-Center(+2.0~+4.9)/ Right(+5.0~+7.9)/ Extreme Right(+8.0~+10)。

事实性评级 = 四维加权(0 最好,10 最差): | 维度 | 权重 | 说明 | |---|---|---| | Failed Fact Checks | 40% | 近五年第三方 fact check 记录,"Only uncorrected errors negatively impact the score";推广伪科学/阴谋论直接 ≥5 | | Sourcing | 25% | 引用与外链可信度与深度 | | Transparency | 25% | 五要素披露:About page / ownership / funding / authors / location | | One-Sidedness / Bias by Omission / Propaganda | 10% | 情绪化语言、单边叙事 |

分档:Very High(0)/ High(0.1–1.9)/ Mostly Factual(2.0–4.4)/ Mixed(4.5–6.4)/ Low(6.5–8.4)/ Very Low(8.5–10)。

MBFC 自我定位(引用时要带的限定语):"no universally accepted scientific formula exists for measuring political bias… ratings are structured human assessments";且 "remains primarily tailored to the political landscape of the United States"(S11)。

3.2 SIFT 四步法(run2 已升 A 级,S16 = hapgood.us 2019-06-19 原文缓存)

run1 抓的是 2019-10-23 错误日期路径(WordPress 软 404,返回页 grep SIFT=0),因此当时只能给 D 级综述。run2 换正确路径 2019/06/19/sift-the-four-moves/,抓到原文(1.15MB,grep SIFT=358 次),四动词逐字命中:

"Stop, Investigate the source, find better coverage, trace the original context." "Trace claims, quotes, and media back to the original context"

四个动作的可训练表述(对应原文):

  1. Stop — "STOP and take a second to remember your purpose. If you just want to repost, read an…" / "stopping periodically and reevaluating our reaction or search strategy is key."(动情绪先停,先确认自己对这条信息的目的)
  2. Investigate the source — "Just as when you hear a rumor you want to know who the source is before reacting…"(先查出处再读内容,lateral reading)
  3. Find better coverage — 原文明示目的是 "to go out and find the…"(更权威的覆盖,而不是死磕单源)
  4. Trace claims, quotes, and media back to the original context(把引文/数据/图追回原始语境)

理论基础是 Stanford History Education Group 的 lateral reading 研究。→ 四动词骨架 + 各自一句话判据可以写进 SKILL 硬判据;但 SKILL 内的「触发阈值/评分权重」若要加,属 C 级自定规范。

3.3 CRAAP Test(仍 C 级,官方页两轮均未落盘)

run1 两处路径 404,run2 换 /help/craap-test 仍 404(58KB 错误页)且 WebFetch 域名校验失败。五准则(Currency / Relevance / Authority / Accuracy / Purpose)与 1–4 打分、45–50 excellent 量表仅经 WebSearch 摘要获得,无逐字原文。 → 只能以「框架名 + 五准则」引用,禁引逐字题项;若 Digester 需要逐字题项,须换镜像源(如各大学 libguides 的转载页)重抓后再升 B 级。

3.4 可落地的信源分级(C 级构造性设计,供 SKILL 采用)

  1. 一手官方(官网文档/官方博客/监管披露/论文)> 2. 政府/标准组织/学术 > 3. 主流媒体(Least Biased 档)> 4. 行业媒体/企业博客 > 5. 聚合与内容农场(不采)。
  2. 每条引用记:URL + 抓取时间 + 哈希 + 抓取方式(本报告 sources_v1.json 即按此格式,可直接当模板)。
  3. 官方教训佐证内容农场问题(A,S1):"human testers noticed that our early agents consistently chose SEO-optimized content farms over authoritative but less highly-ranked sources like academic PDFs or personal blogs. Adding source quality heuristics to our prompts helped resolve this issue."——不加密源启发,agent 会天然偏向 SEO 农场。

§4 交叉验证与事实核查

  • 引用回填由独立环节完成(A,S1):CitationAgent "processes the documents and research report to identify specific locations for citations… ensures all claims are properly attributed to their sources."——「生成」与「标注」分离,值得照搬。
  • 评审判据即核查清单(A,S1): LLM-as-judge 五维 rubric —— factual accuracy("do claims match sources?")、citation accuracy("do the cited sources match the claims?")、completeness、source quality("did it use primary sources over lower-quality secondary sources?")、tool efficiency;且 "a single LLM call with a single prompt outputting scores from 0.0-1.0 and a pass-fail grade was the most consistent"。
  • 逐资源留痕(B,S7): "Summarize and source-track each resource" + "Aggregate over 20 sources for objective conclusions"。
  • 多源取众数降错率(B,S7):见 §1.2 引文。
  • 学术命名的两个编造类缺陷(B,S9):source bias transfer(信源偏差迁移)、over-association of unrelated facts(无关事实过度关联)。
  • 多源冲突仲裁:已缓存信源中无任何系统性官方条款 —— 属本报告如实标注的缺口。C 级建议:按「独立域数量 × MBFC 事实性档位 × 是否一手」做分歧矩阵,一手>多源一致>单源高可信>其余标存疑;此条无官方依据。
  • 诚实留痕机制(C,本报告自证可行):negative_findings(搜了没找到)与 blocked_sources(登录墙/不可达)必须随报告交付,禁止静默省略 —— 本轮 sources_v1.json 即含两区块。

§5 深度调研报告的结构范式与反模式

5.1 结构范式

系统 范式 级别
STORM outline(pre-writing)→ 长文 + 引用 → polish(摘要节 + 去重) B
GPT Researcher 问题集 → 逐问题收集 → 摘要留痕 → publisher 聚合,>2000 词、20+ 源 B
Anthropic Research 检索计划(入 Memory)→ 并行 subagent → 综合 → CitationAgent 回填引用 → 交付 A

C 级建议骨架(融合三者):需求澄清与检索计划 → 子问题×视角矩阵 → 证据表(claim/来源/等级/时间)→ 分级结论(确定/推定/存疑)→ 引用清单 → negative_findings + blocked_sources → 局限声明。

5.2 官方点名的反模式(全部 A 级,S1)

  1. "spawning 50 subagents for simple queries"(简单任务滥开并行)
  2. "scouring the web endlessly for nonexistent sources"(找不到硬找,不停机)
  3. "agents continuing when they already had sufficient results"(够用不停)
  4. "using overly verbose search queries"(查询过长过窄)
  5. 模糊派单 → "agents duplicate work, leave gaps"
  6. 串行搜索:"Our early agents executed sequential searches, which was painfully slow"
  7. SEO 内容农场偏好(§3.4)
  8. 上下文溢出丢计划(200K 截断 → 必须外部 Memory)
  9. "excessive updates"——subagent 之间互相干扰
  10. 产物走协调者转发的「传话游戏」:"Subagent output to a filesystem to minimize the 'game of telephone.'"——大件落盘、只回传轻量引用(对 Catania/LemonBoss 体系的文件交付约定是同型条款)

5.3 与本组织的接合点(A + C)

  • 官方就有 deep-research skill 模板(A,S3 = code.claude.com/docs skills 页): --- name : deep-research description : Research a topic thoroughly context : fork agent : Explore --- Research $ARGUMENTS thoroughly : 1. Find relevant files using Glob and Grep 2. Read and analyze the code 3. Summarize findings with specific file references 这是官方文档给出的字面同型模板(含 context: fork 隔离上下文、子代理回传摘要的机制),可作 SKILL 骨架母本;与 h3-v4 调研中「找官方同型模板钉死母本」的方法论一致。
  • skill 工程硬约束(A,S3):description 截断在 1,536 字符;"Keep the body itself concise. Once a skill loads, its content stays in context across turns, so every line is a recurring token cost."
  • 工具设计(A,S2 "Writing effective tools for AI agents"):工具描述质量直接决定 agent 成败(同一博客体系,S1 亦引 "Bad tool descriptions can send agents down completely wrong paths");工具自测试 agent 改写描述后 "resulted in a 40% decrease in task completion time"。
  • C 级:deep-research Agent 应产出可复用的「检索计划 → 证据表 → 分级结论」三件套,并把每个检索结果按 §3.4 分级打标 —— 与本报告 sources_v1.json 的 schema 保持一致即可直接复用。
  • 运行时定位(A,S5 = anthropics/claude-code README):"Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks… all through natural language commands."——deep-research skill 落在 Claude Code 的 agent loop 上跑(与 S3 的 context: fork 机制同一体系),不是独立服务。

§6 网络受限记录(SOP ④,如实登记)

主机 症状 复检次数
openai.com HTTP 403(Cloudflare 拦截页);WebFetch 被企业策略拦截 curl 1 + WebFetch 1
www.perplexity.ai TCP connect timeout 2
support.google.com timeout 45s/40s 3
en.wikipedia.org TCP connect timeout 2(WebFetch 1)
webliteracy.pressbooks.com DNS 解析失败 1
hapgood.us HTTP 404(WordPress 软 404,返回页 grep SIFT=0) 1
library.csuchico.edu / www.csuchico.edu HTTP 404(CRAAP PDF 已迁走) 2
raw.githubusercontent.com timeout(0 字节) 2 → 经 cdn.jsdelivr.net/gh//@/ 镜像可达,全部 200
docs.claude.com(部分页) HTTP 200 但正文为 "App unavailable in region" 区域封锁页(两条 URL 返回字节级相同内容,sha256 一致) 1
  • 本环境 WebFetch / WebSearch 工具均不可用("Unable to verify if domain ... is safe to fetch" / 无实时检索),全部抓取走 curl。
  • 本轮 SOP 执行自检:R1 首抓即落盘 sources_raw/ 并算 sha256 ✅;R2 curl 带超时与重试参数 ✅;R3 无词级重复现象,未触发;单点失败均按「换路→WebFetch→标 blocked→继续」处置,未重跑整轮 ✅。
  • 遗留给上级的提醒:--retry 3 --retry-all-errors --max-time 60 在慢主机上会把单条 URL 放大到 4 分钟以上(本轮实测先超时),建议 SOP 把这个组合改为 --max-time 40 --retry 1 + 失败换镜像域。

§7 给 Digester 的输入清单(run2 更新)

  1. 骨架母本:§5.3 官方 deep-research skill 模板(A 级)+ STORM 四模块(B 级)+ Anthropic orchestrator-worker 流程(A 级)+ run2 补充的 OpenAI「Search/Fetch 两工具」最小接口(§9.3,A 级)。
  2. 必写进 SKILL 的 A 级硬判据:派单四要素、先宽后窄、effort scaling、停机判据、信源启发、产物落盘轻量回传、外部 Memory 存计划;run2 新增——检索计划先给用户 revise/approve(S15)、逐 claim 可引且引用内联(S12/S14)、同源不重复取件(S14)、过程步骤与已用信源可见(S12)、SIFT 四动词(S16)。
  3. 不可作硬判据的:Perplexity 任何机制(仍 D 级未复检) 与 CRAAP 逐字题项(仍 C 级)。SIFT / OpenAI / Gemini 已由 run2 升级为 A 级,可以用了。
  4. 需上级拍板的裁决点(无官方依据):多源冲突仲裁规则(§4 缺口项,run2 复核维持无官方条款)。

§9 run2 增量复核(2026-08-28 08:15–08:37 UTC,D→A 升级)

本章只写 run2 新增/修正的事实,不重写 run1 已核实的结论(§1–§5 引用位置不变)。 每条都带信源编号(对应 sources_v1.json 的 run: 2 条目)与可复核缓存文件。

9.1 OpenAI Deep Research:官方博客全文(S12,A 级)

openai.com 原站两轮均 403,run2 改抓 OpenAI 官方 openai-cookbook 仓库自带的博客 PDF 镜像(路径由官方仓库 tree API 枚举定位),内容等价。PDF 为文本层解码抽取(Identity-H + ToUnicode 自写解码器),存在系统级字形替换(f→y、d→→、I→k 等),以下引文已按上下文人工修复并逐条可回查缓存 PDF:

  • 定位与产出形态:"a new agentic capability that conducts multi-step research on the internet for complex tasks. It accomplishes in tens of minutes what would take a human many hours."(§1.4 旧表述「通常 5–30 分钟」来自同文后段 "may take anywhere from 5 to 30 minutes",两处官方时长口径一致)
  • 规模:"you give it a prompt, and ChatGPT will find, analyze, and synthesize hundreds of online sources to create a comprehensive report at the level of a research analyst."(「浏览数百网站」这一量级表述的官方归属是 OpenAI 而非 Google,见 NF3)
  • 可溯源性(训练反编造条款的直接官方依据):"Every output is fully documented, with clear citations and a summary of its thinking, making it easy to reference and verify the information."
  • 过程透明:"Once it starts running, a sidebar appears with a summary of the steps taken and sources used."(= 检索步骤与已用信源要随过程可见,不只是最终报告)
  • 与普通对话的分工:"deep research's ability to conduct extensive exploration and cite each claim is the difference between a quick summary and a well-documented, verified answer that can be usable as a work product."(「逐 claim 可引」是把调研结果当工作产物的分界线,可作 SKILL 的产出标准)

9.2 OpenAI Deep Research System Card(S13,A 级)

  • "Deep research is a new agentic capability that conducts multi-step research on the internet for complex tasks. The deep research model is powered by an early version of OpenAI o3 that is optimized for web browsing."
  • "Deep research leverages reasoning to search, interpret, and analyze massive amounts of text, images, and PDFs on the internet, pivoting as needed in reaction to information it encounters."——「随检索所获动态转向」是官方对非线性行为的又一条表述(与 S1 "linear, one-shot pipeline cannot handle these tasks" 同向)。
  • 附加能力:"It can also read files provided by the user and analyze data by writing and executing python code."(输入不只是 URL,可含用户文件与代码分析)
  • 安全侧(若 SKILL 要写部署门槛):Preparedness Scorecard 规则 "Only models with a post-mitigation score of 'medium' or below can be deployed."

9.3 OpenAI 官方对 Deep Research 工具接口的最小约定(S14,A 级)+ Gemini 官方博客(S15,A 级)

S14(官方 Cookbook):

  • "The Deep Research agent relies specifically on Search and Fetch tools. Search should look through your object store for a set of specfic, top-k IDs. Fetch, is a tool that takes objectIds as arguments and pulls back the relevant resources."——「检索找 ID、拉取取全文」两工具分工是官方对最小可用 Deep Research 接口的定义(错别字 specfic 系原文如此)。
  • 官方 system_message 示例可提炼的判据:"Prioritize reliable, up-to-date sources: peer-reviewed research, health organizations…";"If you've already retrieved a file, do not call fetch again for that same file."(同源去重/不重复取件);"Include inline citations and return all source metadata."(引用内联 + 元数据齐全)。

S15(Google 官方博客):

  • "Deep Research uses AI to explore complex topics on your behalf and provide you with findings in a comprehensive, easy-to-read report…"
  • "After you enter your question, it creates a multi-step research plan for you to either revise or approve."(human-in-the-loop 的 A 级依据,已回填 §2.1)
  • 产出物定义:"providing a comprehensive report with key findings and links to original sources."(= 关键发现 + 原始链接,官方对报告最低构成的说法)

9.4 SIFT 原始出处(S16,A 级)与 STORM 评测细节(S17,B 级)

SIFT 四动词逐字引文已直接写进 §3.2,此处只记出处修正:run1 用的 2019-10-23 是错误日期路径(WordPress 软 404),正确路径 2019/06/19/sift-the-four-moves/,1.15MB 缓存、grep SIFT=358 次。

S17(STORM 全文 HTML)补两条可训练的量化判据:

  • 可验证性的操作定义:"For verifiability, we calculate the citation recall and citation precision based on the definition in Gao et al. (2023). We use Mistral 7B-Instruct to examine whether the cited passages entail the generated sentence."——即「引用召回 = 该引没引、引用精度 = 引文是否真支撑该句」,这是交叉验证章节可直接复用的两个指标;STORM 自评 84.83 / 85.18,且原文自己承认 "we acknowledge the possibility of the evaluator LLM overrating machine-generated text"。
  • 报告结构评分维度:官方 1–5 分 rubric 的 Broad Coverage 档位逐字命中(Score 1 "Severely lacking; offers little to no coverage…" → Score 5 "Exemplary in breadth…"),可作「覆盖度评分」的现成模板。

9.5 run2 补充信源与网络窗口差异

  • 新增缓存(run: 2,共 8 条有效 + 2 条辅助件):S12 OpenAI 博客 PDF、S13 系统卡、S14 MCP README、S15 Gemini 博客、S16 SIFT 原文、S17 STORM 全文 HTML、S18 Open Deep Research README、S19 Full Fact toolkit、S20/S21 两个 tree API 辅助件。
  • S18(LangChain 官方开源实现)补一条可训练流水线:四段 Summarization / Research / Compression / Final Report 逐字命中,且公开实测成本(100 题 $45.98–$187.09、RACE 0.4309–0.4943)——「深度调研很贵」有数字可引,effort scaling 判据可据此定预算档位。
  • S19(Full Fact)限缩使用:是索引导航页(纯文本 5.2KB),只证明「分主题核查指南存在」,不含逐步方法论文本(NF4)。
  • 网络窗口差异(对 SOP 的直接输入):run1 时 raw.githubusercontent.com 不可达、只能走 jsdelivr 镜像;run2 同一主机直连全部 200。说明本环境网络出口是时变的,SOP 应写成「先直连 → 失败换镜像域 → 再失败记 blocked」,而不是把某一台主机永久标黑。run1 提的 --retry 3 --retry-all-errors --max-time 60 会把单条 URL 放大到 4 分钟以上的问题在 run2 复现(docs.perplexity.ai 两次 rc=28 耗尽全部重试),该建议维持。
  • run2 的 WebFetch 对 openai.com / library.csuchico.edu / docs.perplexity.ai 仍报 "Unable to verify if domain is safe to fetch",与 run1 结论一致;但 WebSearch 在 run2 可用(CRAAP 五准则即由此部分恢复)——run1 记录的「WebSearch 不可用」是当时的网络状态,不是本环境常量。

9.6 两轮合并后仍未闭合的缺口(留给 Digester / 下一轮)

  1. Perplexity Pro Search:D 级,两轮不可达,唯一仍缺的主流系统一手材料。
  2. CRAAP 逐字题项:C 级,官方页两轮 404,需换镜像源。
  3. 多源冲突仲裁:无任何官方条款(NF6),仍是最需要上级拍板的裁决点。
  4. 报告结构标准:确认不存在官方标准(NF5),SKILL 里的结构骨架必须标 C 级自定规范。

§8 run1 未尝试抓取的信源(供后续轮次)

  • docs.gptr.dev(GPT Researcher 详细文档)、storm-project.stanford.edu、aclanthology.org(Co-STORM EMNLP 版)、agentskills.io、support.claude.com 的三个 skills 条目(S4 内链)。
  • 需要时直接抓,STORM 论文细节可直接读已缓存全文 PDF(S10,27 页,1.3MB)。