情报台一号 · 架构研究库Desk 1 · Architecture Research Library
AI 架构研究库The AI Architecture Research Library
作者 深翼DeepWing｜维护 情报台一号｜最后核查 2026-10-29By DeepWing | Maintained by Desk 1 | Last checked 2026-10-29
把真实运行的 AI 系统逐层拆开，只写有出处的部分：哪一层用了什么，付出了什么代价，还有什么查不到。查不到的如实标未知，不猜，不补。这本库不拼收录数量，拼的是每一条判断都能被读者自己复核。Real, running AI systems, taken apart layer by layer. Only what sources support is written as fact: what each layer does, what it costs, and what cannot be found out. What cannot be verified is marked unknown, in the open. Nothing is guessed and nothing is padded. This library does not compete on how many systems it lists; it competes on whether every judgement in it can be checked by the reader.
分析坐标：七层与十关（DW-AI-ARCH v1.0）Coordinates: the seven layers and ten gates (DW-AI-ARCH v1.0)
证据口径：四档等级，逐条可追溯Evidence: four grades, every load-bearing claim traceable
当前状态：首批深拆解三条，两条已定版（ARC-012、ARC-013），一条在建（ARC-011）；登记清单见图谱页Status: three first deep teardowns: two published (ARC-012, ARC-013), one in progress (ARC-011); the register lives on the atlas page
来这里，能回答什么What you can answer here
问一，理解系统：这个系统实际用了哪些模型、编排、工具与控制措施？哪些是官方确认的，哪些仍然查不到？One, understand a system: Which models, orchestration, tools and controls does this system actually use? What is officially confirmed, and what still cannot be found out?
问二，比较设计：两个系统处理权限、记忆与失败恢复的方式有何不同？各自付出了什么代价？Two, compare designs: How do two systems differ in handling access control, memory and failure recovery, and what does each choice cost?
问三，做出技术判断：要自建或采购同类系统，哪些实现值得参考，哪些风险必须先验证？现有证据够不够支撑这个决定？Three, make a technical judgement: If you were building or buying a similar system, which implementations are worth borrowing, which risks must be verified first, and is the current evidence strong enough to support the decision?
这本库怎么读How to read this library
七层是坐标，不是分层图：模型、编排、工具、记忆、知识、门禁、盯梢，是深翼的标准化分析坐标，用同一套语言拆所有系统，不要求真实系统按这七格分家。一项机制可以落在多层，条目里定一个主要归属，其余层只作指向，不重复描述。The seven layers are coordinates, not a stack diagram: Model, Orchestration, Tool Use, Memory, Knowledge, Access Control and Observability are DeepWing's standard analytical coordinates, one shared language for taking any system apart. Real systems are not required to divide themselves along these lines. A mechanism may map to several layers; each entry names one primary home and points from the others instead of repeating the description.
每层四栏：已证实的实现、未知、设计取舍、来源，四栏分开写。深翼的推断只进设计取舍一栏，并标明「深翼推断」，不混进已证实的实现。Four columns per layer: verified implementation, unknowns, design trade-offs, and sources, written separately. DeepWing inferences appear only in the trade-offs column, labelled as DeepWing inference, never inside verified implementation.
证据等级四档：Four evidence grades:
一手已核：直读官方文档、论文原文、代码库或监管文件核对过。
多源已核：两个以上独立来源相互对得上，其中至少一源接近一手。
媒体转述：只有媒体报道，未见一手原文，按转述记，不作承重结论。
待核：已登记，尚未核查，暂不引用。
Primary-verified: checked against official documentation, the original paper, the code repository or a regulatory filing, read first-hand.
Multi-source verified: two or more independent sources agree, at least one of them close to primary.
Media-reported, pending: media coverage only, no primary text seen; recorded as reporting, never as a load-bearing conclusion.
Unverified: registered but not yet checked; not cited.
条目的证据等级取全条承重事实里最弱的一档。等级只表示来源核实到什么程度，不表示该系统经过独立评审、监管认证或生产验证。An entry's grade is the weakest grade among its load-bearing facts. A grade states how far the sourcing has been verified; it does not mean the system has passed independent review, regulatory certification or production validation.
未知如实标注：公开材料没有披露的，逐项写明未知什么、为什么未知，并计入条目头的未知计数。未知不是缺漏，是这本库的正式数据。Unknowns are labelled, not hidden: where public material does not disclose something, the entry states what is unknown and why, and counts it in the unknown count in the entry header. An unknown is not a gap in the writing; it is part of the data.
关联三状态：条目与模式、判例、自检项之间的关联分三档记：已确认关联；已检查，未发现足够证据；尚未检查。查过没有，就记没有。无关联可以是一条有效的研究结论，不为凑数挂链。Three states for links: links between an entry and patterns, cases and self-check items are recorded in three states: confirmed; checked, with insufficient evidence found; and not yet checked. If the check found nothing, the entry says so. Having no link can be a valid research finding, and links are never added to fill a quota.
如何像架构师一样思考How to think like an architect
系统思维Think in systems
看见连接，识别规律，找到杠杆点。See the connections, spot the patterns, find the leverage points.
为变化而设计Design for change
选能弯曲的结构；今天固定的东西，明天就会变。Choose structures that can flex. What is fixed today will change tomorrow.
权衡分析Trade-off analysis
没有完美方案。比较选项，刻意取舍。There are no perfect solutions. Compare the options and choose deliberately.
场景思维Scenario thinking
预演几种未来，把可能性和风险提前看清。Play out several futures to surface possibilities and risks early.
复用模式Reuse patterns
成熟的蓝图省时间；也要知道什么时候打破定式。Proven blueprints save time. Know when to break the mold.
长期主义Think long-term
速成方案会留下债。不只看今天，也看明天。Quick fixes leave debt behind. Weigh tomorrow, not just today.
以用户为中心Customer-centric perspective
解决真实问题，不做漂亮的抽象。Solve real problems, not elegant abstractions.
影响与说服Influence and persuasion
好想法需要认同。把「为什么」讲清楚。Good ideas need buy-in. Make the why clear.
首批深拆解First deep teardowns
第一批按统一模板重做的三个真实系统，正在施工。完工一条，上线一条；上线之前，有现成登记的先链到图谱页的登记，尚无登记的一条先链到图谱页总览，并如实注明待立册。The first three real systems being rebuilt under the unified template are in progress. Each one goes live as it is finished; until then, each card links to its current register entry on the atlas page where one exists, and to the atlas overview where an entry has not been posted yet, saying so plainly.
一条代码变更如何过七道环节：作者端自验、AI 审查、安全审查、自动化检查、人工批准、合并队列、回滚与事后监测兜底。本条的一手材料最厚，用作全库的模板验证条。How one code change passes seven stages: author self-checks, AI review, security review, automated checks, human approval by risk tier, the merge queue, and rollback with post-merge monitoring as the backstop. This entry carries the thickest primary material and serves as the template validation entry for the whole library.
状态：v1.0 已定版（2026-10-27 核查）Status: v1.0 published (checked Oct 27, 2026)
冻结基座模型、只改脚手架文件的自改进框架。按统一模板逐层写完：一手已核（论文与代码仓双来源），承重判断 51 条、未知 10 项；判例侧确认关联 INC-030。本条的门禁结论：护栏本身在可编辑清单内，护栏被改弱时的单独复核未见披露。A self-improving framework that freezes the base model and edits only the scaffold files. Written layer by layer to the shared template: primary-source verified against the paper and the code repository, 51 load-bearing claims with 10 unknowns, and a confirmed link to case INC-030. The gate finding: the guardrails file is itself editable, and no separate review of a weakened guardrail is disclosed.
状态：v1.3 已定版（2026-10-28 核查，2026-10-29 修订）Status: v1.3 published (checked Oct 28, 2026; revised Oct 29, 2026)
跨层权限控制、审计与生命周期治理。按统一模板逐层写完：一手已核（三篇官方博客加文档站补核），承重判断 62 条、未知 12 项；判例侧确认关联 INC-014、INC-015、INC-033、INC-038。本条的门禁结论：凭证永不进沙箱，出站凭证由网关验证三项身份后才换入；残余缺口在智能体独立身份的下线收尾，以及网关策略的临时豁免与误配回滚。Cross-layer access control, audit and lifecycle governance. Written layer by layer to the shared template: primary-source verified against three official blog posts plus a documentation check, 62 load-bearing claims with 12 unknowns, and confirmed links to cases INC-014, INC-015, INC-033 and INC-038. The gate finding: credentials never enter the sandbox, and the gateway swaps in outbound credentials only after checking three identities; the remaining gaps are in retiring an agent’s own identity, and in exemptions and rollback for gateway policies.
本页是门面与导读，不重复造表。架构、产品拆解与新思想模式的登记清单在图谱页：按十关筛选，逐条带编号、所属关与来源。This page is the front door and the reading guide; it does not duplicate the tables. The register of architectures, product teardowns and idea patterns lives on the atlas page, filterable by the ten gates, every entry carrying its ID, its gates and its sources.
判例库：架构上的决定最后如何失效，看判例库里的真实事故。每条深拆解都与相关判例互链，失效模式逐条挂得回事故。The case library: how architectural decisions eventually fail, in real incidents. Every deep teardown links both ways with the cases behind its failure modes.
企业 AI 架构自检：把在本库读到的检查项，拿回自己的系统上验一遍。17 题，10 个检查点，三分钟出报告。The enterprise AI architecture self-check: take the check items you just read about and run them against your own system. 17 questions, 10 checkpoints, a report in about three minutes.
框架规范 DW-AI-ARCH v1.0：七层与十关的定义、条文与检查项全名在本规范里。正式版已存证于 Zenodo，DOI：10.5281/zenodo.23715226。The framework, DW-AI-ARCH v1.0: the definitions, clauses and full check-item names for the seven layers and ten gates. The formal edition is deposited on Zenodo, DOI: 10.5281/zenodo.23715226.
方法与独立性：本库由深翼 DeepWing 独立维护，不接受被研究厂商的赞助。厂商自报的数字一律归属标注，不作本库结论。记错了公开更正，旧说法留痕。（「独立性」全文见 /about/ 关于页。）Method and independence: this library is maintained independently by DeepWing and takes no sponsorship from the vendors it studies. Vendor-reported figures are always attributed and never presented as this library's conclusions. Errors are corrected in public, with the earlier wording kept on record. (Corrections link to /corrections/; the full independence statement is on the About page, /about/.)
